<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Developer tools - Soursen</title><description>Developer tools covers the CLIs and APIs engineers run, including what a tool collects and where that data goes, from the docs or the release note.</description><link>https://soursen.org/</link><language>en-gb</language><atom:link href="https://soursen.org/beat/developer-tools/rss.xml" rel="self" type="application/rss+xml"/><item><title>Adobe Commerce Upgrade Reportedly Cut a $6 Million Brand’s Sales 30% on Launch Day</title><link>https://soursen.org/articles/adobe-commerce-upgrade-reportedly-cut-a-6-million-brands-sales-30-on/</link><guid isPermaLink="true">https://soursen.org/articles/adobe-commerce-upgrade-reportedly-cut-a-6-million-brands-sales-30-on/</guid><description>Founder Briony Kennedy says a website migration cut sales by about 30% and broke roughly $1 million a year in ads. Adobe will not comment. The $150 million Adobe settlement in the headlines concerns consumer subscriptions and has no stated link to her contract.</description><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On November 8, 2026, 7NEWS Australia reported that Adorn Cosmetics, a Melbourne skincare company that once turned over about $6 million a year, was about to close. Founder Briony Kennedy told the network her troubles began when she moved her online store to Adobe’s ecommerce platform. She laid off staff, sold her family home, rebuilt the store on Shopify, and says it came too late. I want to walk through what she claims, what the public record backs up, and what remains unknown, because this story is being passed around as a lesson about Adobe when the evidence supports a narrower and more useful lesson about launching a store.&lt;/p&gt;
&lt;h2 id=&quot;what-briony-kennedy-says-happened&quot;&gt;What Briony Kennedy says happened&lt;/h2&gt;
&lt;p&gt;Adorn sells refillable, recyclable and biodegradable skincare, mostly to mature aged women, and Kennedy started it in 2009 as a new mother of twins, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://business.adobe.com/au/blog/basics/adorn-cosmetics-gives-its-e-commerce-experience-a-makeover-with-adobe-commerce&quot;&gt;Adobe’s own case study of the company&lt;/a&gt;. Her account to 7NEWS has 5 parts. Adobe convinced her to leave a basic free website for a paid ecommerce platform. Adobe quoted revenue increases between 34% and 94%. A Melbourne agency named Convert Digital did the implementation. When the new site launched, sales fell by about 30%, and roughly $1 million a year of Instagram and Facebook advertising stopped sending shoppers to the website.&lt;/p&gt;
&lt;p&gt;She says she was not allowed out of the contract. She says she wanted the move for security and expected it to lift sales. After the family home was sold and the staff were let go, she abandoned the Adobe site, started again on Shopify, and was left with debt and a looming bankruptcy. Her closing advice to other owners was short, and I think it is the most quotable line in the report. She told viewers “Don’t chase the new shiny thing.” &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://x.com/7NewsMelbourne/status/2119338248723992783&quot;&gt;The 7NEWS Melbourne post&lt;/a&gt; summarizes the story in 2 sentences, and every figure in it comes from Kennedy herself.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DGnqfOK4_26SJHM.webp&quot; alt=&quot;The projected revenue gain Adobe is said to have quoted, next to the sales change Kennedy reports at launch.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The projected revenue gain Adobe is said to have quoted, next to the sales change Kennedy reports at launch. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://x.com/7NewsMelbourne/status/2119338248723992783&quot;&gt;7NEWS Australia, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;adobe-and-the-agency-have-said-almost-nothing&quot;&gt;Adobe and the agency have said almost nothing&lt;/h2&gt;
&lt;p&gt;An Adobe spokesperson told 7NEWS the company would not comment on the case for legal reasons, and the report says a mediation attempt between the parties failed. Convert Digital is named in the report as the implementation partner, and I found no public statement from the agency in any coverage I read. That leaves a one sided record. The report gives no dates for the contract, the launch or the move to Shopify. It gives no baseline or time window for the 30% drop. It does not say whether a formal insolvency process has started, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ppc.land/adorn-cosmetics-faces-closure-blaming-30-sales-drop-on-adobe-site/&quot;&gt;PPC Land’s analysis of the report&lt;/a&gt; lists the sales decline, the $6 million turnover and the $1 million ad spend among the figures with no supporting data in public.&lt;/p&gt;
&lt;p&gt;None of that makes Kennedy wrong. A founder describing her own losses is a primary source for her own experience. It does mean every number in this story is a claim, and I have treated each one that way.&lt;/p&gt;
&lt;h2 id=&quot;the-platform-she-started-on-was-already-free&quot;&gt;The platform she started on was already free&lt;/h2&gt;
&lt;p&gt;The detail that changed how I read the story is the starting point. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bandt.com.au/adorn-cosmetics-gets-a-digital-makeover-using-adobe-commerce/&quot;&gt;B&amp;amp;T reported in September 2021&lt;/a&gt; that Adorn had used an open source version of Magento since launch, and was upgrading to Adobe Commerce. Adobe bought Magento in 2018 for a reported $1.68 billion, and the free edition is now called Magento Open Source. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dotdigital.com/case-studies/adorn-cosmetics/&quot;&gt;Dotdigital’s case study&lt;/a&gt; says Adorn moved to Magento 2, now branded Adobe Commerce, in 2021, and built a stack around it with Dotdigital for email and Yotpo for reviews.&lt;/p&gt;
&lt;p&gt;Adobe’s case study listed what the upgrade was meant to deliver. Search across ingredients and accreditations, a loyalty program beyond a simple spend and reward plugin, stock levels that sync with inventory, and PCI compliance with security scanning. It cited about 75,000 Facebook followers and 20,000 Instagram followers, and B&amp;amp;T put the social following near 100,000. Those are real assets for a brand selling direct to shoppers.&lt;/p&gt;
&lt;p&gt;Adobe does not publish Adobe Commerce prices. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mgt-commerce.com/blog/magento-license-cost/&quot;&gt;MGT Commerce’s 2026 breakdown&lt;/a&gt; estimates licenses starting around $22,000 a year for the on premise edition and around $40,000 a year on Adobe’s cloud, rising with sales volume. Open Source costs nothing in license fees, and free software still needs hosting, development and security patching. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://scandiweb.com/blog/magento-pricing/&quot;&gt;scandiweb’s cost guide&lt;/a&gt; offers planning anchors and says they are not quotes. It puts a straightforward Open Source build at $20,000 to $60,000, an Adobe Commerce initial build at $60,000 to $500,000 or more, and annual maintenance at roughly $4,000 to $25,000 for Open Source and $15,000 to $50,000 for Adobe Commerce.&lt;/p&gt;
&lt;p&gt;Those are ranges from public guides, and nobody has published Adorn’s actual bill. They do show that moving up an edition adds a license fee and a project budget to a business that was already running.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CzCZm8X2_2qSge4.webp&quot; alt=&quot;Estimated annual license floor for 3 Magento editions. Adobe does not publish its prices.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Estimated annual license floor for 3 Magento editions. Adobe does not publish its prices. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mgt-commerce.com/blog/magento-license-cost/&quot;&gt;MGT Commerce, Mar 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-34-to-94-promise-has-no-public-source-i-can-find&quot;&gt;The 34% to 94% promise has no public source I can find&lt;/h2&gt;
&lt;p&gt;The revenue range is the claim I most wanted to verify. I could not find 34% to 94% in Adobe’s case study or in the B&amp;amp;T report, and PPC Land reached the same result. It notes that an Adobe sponsored IDC study listed in December 2021 reported 29% higher platform revenue for Adobe Commerce users, which is a different number from a different kind of document. Kennedy may have received the range in a sales pitch or a proposal, and a private document would not show up in public searches. Until someone publishes it, the range stays an attributed claim.&lt;/p&gt;
&lt;p&gt;Sales projections like this deserve scrutiny in any platform deal. Ranges like this typically summarize past customers, and they say nothing about what happens to a particular store in its first month after launch. A store that loses 30% at launch and a store that gains 94% a year later can both appear in the same average.&lt;/p&gt;
&lt;h2 id=&quot;how-a-launch-can-break-1-million-of-ads&quot;&gt;How a launch can break $1 million of ads&lt;/h2&gt;
&lt;p&gt;Kennedy’s most specific claim is that her ads stopped linking to the website. The report does not say why, and PPC Land lists the usual suspects without confirming any of them. Product URLs may have changed. Product feed links may have gone out of date. Tracking may have broken. I can tell you how that failure usually works, as long as you read it as a general mechanism and not a finding about Adorn.&lt;/p&gt;
&lt;p&gt;Every ad is a link to a specific address on your site. A platform move often changes those addresses, because the new system builds product and category paths its own way. If the old addresses are not mapped to the new ones, each ad click lands on an error page, and what happens next depends on the ad platform. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://support.google.com/adspolicy/answer/16428019&quot;&gt;Google Ads’ Destination not working policy&lt;/a&gt; disapproves ads whose landing page returns an error such as a 404, which stops the spend and the sales together. I could not confirm how Meta’s systems handle the same failure, and Kennedy’s ads ran on Instagram and Facebook. She puts her ad spend at about $1 million a year, which works out to roughly $2,740 a day and about 17% of her reported $6 million peak turnover. If a platform kept charging for clicks that hit error pages, a week of that would cost about $19,000, before counting the sales that never happened.&lt;/p&gt;
&lt;p&gt;The same mapping problem hits search traffic. Google’s guidance for site moves says to set up a permanent 301 redirect &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.searchenginejournal.com/google-keep-301-redirects-in-place-for-a-year/&quot;&gt;from every individual page on the old site to its equivalent on the new one&lt;/a&gt;, and to keep those redirects in place for at least a year. Google’s site move guide also says &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ppc.land/googles-site-move-guide-now-covers-www-and-non-www-domain-variants/&quot;&gt;permanent redirects do not cause a loss in PageRank&lt;/a&gt;. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.searchenginejournal.com/ask-an-seo-how-can-we-recover-drop-in-organic-traffic-from-site-migration/550915&quot;&gt;Search Engine Journal’s Ask an SEO column&lt;/a&gt; calls missing or incorrect redirects one of the top causes of traffic loss after a migration.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Caw8arsy_2toOdQ.webp&quot; alt=&quot;Kennedy’s reported peak turnover and annual social ad spend. The 2 figures may come from different years.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Kennedy’s reported peak turnover and annual social ad spend. The 2 figures may come from different years. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ppc.land/adorn-cosmetics-faces-closure-blaming-30-sales-drop-on-adobe-site/&quot;&gt;7NEWS Australia via PPC Land, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.practicalecommerce.com/?p=58248&quot;&gt;Practical Ecommerce’s Jill Kocher Brown&lt;/a&gt; documented 2 cases that show the range. One consumer goods site changed every URL in a redesign, used a 301 redirect strategy, lost 25% of its traffic for about 6 weeks, and later ran 19% above its pre launch level. Another manufacturer changed URLs without redirects and watched organic search traffic fall 42% after 12 days. Adorn’s reported 30% sales drop sits between those 2 outcomes, which fits a launch that went wrong without proving that redirects were the cause.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.D4M2k1Kf_1bjsl9.webp&quot; alt=&quot;Reported changes after a platform move. The Adorn figure is sales, and the other 2 are organic search traffic.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Reported changes after a platform move. The Adorn figure is sales, and the other 2 are organic search traffic. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.practicalecommerce.com/?p=58248&quot;&gt;Practical Ecommerce, 7NEWS Australia&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;If the 30% applied to a $6 million business, that is about $1.8 million a year in lost sales. A business can absorb a month of that, and a year of it is far harder, and the report says the damage ended with Kennedy selling her family home.&lt;/p&gt;
&lt;h2 id=&quot;the-150-million-settlement-is-a-different-case&quot;&gt;The $150 million settlement is a different case&lt;/h2&gt;
&lt;p&gt;The 7NEWS report mentions that the US government sued Adobe in June 2024 and that this year Adobe reached a $150 million settlement without admitting wrongdoing. Viewers could easily read that as the legal consequence of what happened to Adorn. It is not described that way anywhere I can find.&lt;/p&gt;
&lt;p&gt;The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.justice.gov/opa/pr/adobe-agrees-150-million-settlement-and-injunction-resolve-alleged-violations-restore-online&quot;&gt;US Department of Justice announced the proposed order&lt;/a&gt; on March 13, 2026. It requires Adobe to pay $75 million in civil penalties and provide $75 million in free services to qualifying customers, and it resolves claims against Adobe and 2 employees under the Restore Online Shoppers’ Confidence Act. The case was about consumer subscriptions, specifically an early termination fee on the “annual paid monthly” plan, which the FTC alleged in 2024 equals 50% of the remaining monthly payments when someone cancels in the first year. The order requires Adobe to disclose that fee before enrollment and to offer simple ways to cancel. Adobe disagrees with the government’s claims and denies wrongdoing, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.pymnts.com/legal/2026/adobe-ends-federal-subscription-lawsuit-with-150-million-settlement-package&quot;&gt;per the settlement reporting&lt;/a&gt;. The Justice Department page was last updated October 20, 2026, and I found no confirmation that the court has entered the order, so I describe it as proposed.&lt;/p&gt;
&lt;p&gt;Nothing in that case concerns Adobe Commerce, merchant contracts or Adorn. The overlap is the word Adobe and a shared theme of exit terms, and I would keep the 2 stories apart until someone shows a connection.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.iNSstHEK_Z20n1cp.webp&quot; alt=&quot;The dates in the Adorn story, with the 2 Adobe settlement items concerning consumer subscriptions.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;698&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The dates in the Adorn story, with the 2 Adobe settlement items concerning consumer subscriptions. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.justice.gov/opa/pr/adobe-agrees-150-million-settlement-and-injunction-resolve-alleged-violations-restore-online&quot;&gt;B&amp;amp;T, US Department of Justice, 7NEWS, PPC Land&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;responsibility-is-still-an-open-question&quot;&gt;Responsibility is still an open question&lt;/h2&gt;
&lt;p&gt;The reaction online split in useful ways. SEO specialist Lily Ray wrote on X that the story is a reminder to hire SEO teams with substantial migration experience before moving a company website, especially an ecommerce one. Nick LeRoy replied that he felt for Kennedy but noted the report points to a migration partner that failed at launch, which raises the question of why Adobe is cast as the villain. Hung Lee wrote that technical people will ask why she chose Adobe, while a non technical owner may not know enough to run the due diligence. These are posts from observers, and none of them has seen the contract.&lt;/p&gt;
&lt;p&gt;Gymshark is another brand that has described leaving Adobe Commerce, and its account comes with a caveat. Founder Ben Francis told Shopify, in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.shopify.com/au/case-studies/gymshark&quot;&gt;a case study on Shopify’s own site&lt;/a&gt;, that building his Adobe Commerce store took 6 to 8 months, that it went down for 8 hours on Black Friday at an estimated cost of £100,000, and that he replatformed to Shopify Plus 10 months into the relationship. Shopify has every reason to tell that story, and the cost is Francis’s own estimate. His remark about customers stuck with me, because he says shoppers blamed the Gymshark brand whichever vendor was at fault.&lt;/p&gt;
&lt;p&gt;I find the timeline the hardest part to explain. The upgrade was announced in September 2021, and the closure is reported in November 2026. The report does not say when the new site launched, how long it ran, or how many months of damage preceded the move to Shopify. Without those dates nobody outside the dispute can say how much of the collapse traces to the launch and how much to other pressures on the business.&lt;/p&gt;
&lt;p&gt;That is also why I think the useful reading is narrow. A platform choice is a decision with a license fee, an implementation partner and a contract. The part that failed, on Kennedy’s account, was the launch itself, and a launch has checks that anyone can demand in writing.&lt;/p&gt;
&lt;h2 id=&quot;2-things-from-this-generalize&quot;&gt;2 things from this generalize&lt;/h2&gt;
&lt;p&gt;First, treat the URL list as a contract deliverable. Before launch, export every address that earns you money, meaning the landing pages of your paid ads, the links in your email templates, your product feed and your top 200 organic pages. Have the agency prove on a staging site that each one either loads or redirects with a single 301 to the right new page, and write “0 broken money URLs at launch” into the acceptance criteria. Then keep the old redirects live for at least a year, as Google advises. Google’s site move guide also says to make changes one after another, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://eseospace.com/blog/geo-during-a-website-redesign-or-migration/&quot;&gt;eSEOspace quotes its example&lt;/a&gt; of moving to a new domain first and changing the layout afterward. A new platform, a new design and new URLs on the same day multiply the ways a launch can fail.&lt;/p&gt;
&lt;p&gt;Second, read the exit clause and set a rollback rule before you sign. Kennedy says she could not leave the contract after the launch went wrong. Ask what it costs to cancel in month 1, month 6 and month 12, and keep the old store able to take orders until the new one has matched its sales for a full 30 days. If the new platform cannot clear that bar, you need the contract to let you go back.&lt;/p&gt;
&lt;p&gt;If you run a store, take 30 minutes this week to export your top 200 landing URLs from your ad accounts and analytics, then check what each one returns today. Any that end in an error page are costing you money before you ever replatform.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>adobe</category><category>marketing</category><category>ai</category><category>ai-agent</category><category>llm</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Andrej Karpathy Rebuilt YouTube in a Weekend Because He Hated the Recommendations.</title><link>https://soursen.org/articles/andrej-karpathy-rebuilt-youtube-in-a-weekend-because-he-hated-the/</link><guid isPermaLink="true">https://soursen.org/articles/andrej-karpathy-rebuilt-youtube-in-a-weekend-because-he-hated-the/</guid><description>On 31 October 2026 one of the people who helped build modern AI put up a YouTube clone sourced from 20 years of Hacker News comments. He wrote almost none of the code himself. The build is a small thing. What it says about who can ship software now is not.</description><pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On a Friday night at 11pm, Andrej Karpathy opened YouTube, got tired of what he called an enshittified wall of slop, and by Saturday morning had a working replacement live on the internet. He called it &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://hackertube.org/&quot;&gt;HackerTube&lt;/a&gt;, and he wrote about it in a short &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;blog post&lt;/a&gt; on 31 October 2026. I want to walk through what he actually built and how, because the what is a neat weekend hack, and the how is the part that should make every working engineer pay attention.&lt;/p&gt;
&lt;p&gt;A quick note on who is talking, since it changes how much weight the claim carries. Karpathy was a founding member of OpenAI, ran AI and Autopilot Vision at Tesla from 2017 to 2022, and in February 2025 coined the term vibe coding to describe building software by typing prompts instead of writing code. When this particular person says he no longer writes code by hand, it is worth taking seriously.&lt;/p&gt;
&lt;h2 id=&quot;what-hackertube-actually-is&quot;&gt;What HackerTube actually is&lt;/h2&gt;
&lt;p&gt;The idea is simple. Instead of trusting YouTube’s algorithm, which optimizes for the median viewer and serves exaggerated thumbnails and the same tired topics, Karpathy pulled video recommendations from the one place he trusts for taste, Hacker News comments. The tool downloaded the full &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://huggingface.co/datasets/open-index/hacker-news&quot;&gt;Hacker News comment dataset&lt;/a&gt;, filtered it, and kept only comments that recommended a YouTube video.&lt;/p&gt;
&lt;p&gt;The funnel is worth seeing in numbers. The filtering narrowed roughly 40,000 recommending comments down to 6,454 distinct YouTube videos, posted and praised by Hacker News users between 2007 and 2026, sorted into 150 topics. Those videos got wrapped in a familiar YouTube style interface, pushed to a server, and secured. The result is a video site where the recommendation engine is two decades of a technical community’s collective judgment rather than a model tuned to maximize engagement.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.BGi01G0J_1EIyh6.webp&quot; alt=&quot;How 20 years of comments became a video library. Each step narrows the set toward videos real people vouched for.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How 20 years of comments became a video library. Each step narrows the set toward videos real people vouched for. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;Karpathy,&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://karpathy.bearblog.dev/&quot;&gt;karpathy.bearblog.dev&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The feature list is small and deliberate. HackerTube shows a frame from 50% into each video instead of the uploaded thumbnail, which kills clickbait at the source, because you see what the middle of the video actually looks like. It builds a personalized ForYou feed with no account and no login, storing state in the local browser session. It has keyboard shortcuts. It also lets you export the whole site’s data so you can hand it to your own agent and build your own version. That last feature is a tell about where Karpathy thinks this is going, software you fork by describing it, not by cloning a repository.&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-matters-he-barely-touched-the-code&quot;&gt;The part that matters, he barely touched the code&lt;/h2&gt;
&lt;p&gt;Here is the sentence from his post that I keep coming back to. Karpathy says he does not do anything resembling traditional engineering anymore, no reading or writing code, no running commands in a terminal. He talks to Claude and has it do everything.&lt;/p&gt;
&lt;p&gt;Look at what counted as manual work on this project. He created a Google Cloud API key through a web UI so the tool could query YouTube for video metadata. He booted up a server. He bought the domain name. That is the entire list of things a human did with their hands. Once the box was running, he gave the model its IP address and told it to update, deploy, secure, harden, and optimize the code, and it did the whole thing. He describes working on other things in between the model’s work, so it still felt a little like teamwork.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CFIGluqo_ZbbPG0.webp&quot; alt=&quot;The split of work on HackerTube. The human contribution was three setup steps outside the code itself.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The split of work on HackerTube. The human contribution was three setup steps outside the code itself. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;Karpathy,&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://karpathy.bearblog.dev/&quot;&gt;karpathy.bearblog.dev&lt;/a&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;He built the tool with Claude Code in the desktop app on a Mac, running Opus 5.5 on medium effort as his default. The spam filtering ran Sonnet over the comment dataset inside a Claude Code Workflow to strip out junk. The whole build, he says, now takes about 2 hours.&lt;/p&gt;
&lt;h2 id=&quot;why-2-hours-is-the-number-to-stare-at&quot;&gt;Why 2 hours is the number to stare at&lt;/h2&gt;
&lt;p&gt;Strip away the YouTube angle and what you have is a real data pipeline feeding a real web application, deployed and hardened on a live server. Downloading and cleaning a large public dataset, running a model over tens of thousands of records to classify them and strip the spam, building a frontend with a recommendation system and session state, provisioning infrastructure, and securing it. A few years ago that is a small team working for a week or two. Karpathy did it solo, mostly by describing what he wanted, in a morning.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Dpax4UUQ_O0zkD.webp&quot; alt=&quot;A directional comparison of build time. The point is the order of magnitude it collapsed by.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;A directional comparison of build time. The point is the order of magnitude it collapsed by. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://karpathy.bearblog.dev/&quot;&gt;Karpathy, karpathy.bearblog.dev, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;This matters beyond one famous person flexing, because the bottleneck moved. The hard part of software used to be the implementation, turning an idea into working code. For a growing class of projects, the implementation is close to free now, and the scarce input is knowing what to ask for and being able to judge whether the result is good. Karpathy still made every real decision here, the data source, the 50% thumbnail trick, the design with no accounts, the export feature. The model did the typing.&lt;/p&gt;
&lt;h2 id=&quot;where-the-easy-case-ends&quot;&gt;Where the easy case ends&lt;/h2&gt;
&lt;p&gt;I want to be careful not to oversell it. HackerTube is a read only site built on a public dataset with no users to protect, no payments, no regulatory surface, and no legacy codebase to avoid breaking. That is close to the easiest possible case for an agent to handle end to end. A banking backend or a 10 year old codebase with 2,000 interlocking tests is a different problem, and anyone telling you those are also 2 hour jobs is selling something.&lt;/p&gt;
&lt;p&gt;What it does prove is that the easy case has gotten very easy, and the set of projects that fall into the easy case keeps growing. The person demonstrating it is not a hype account. He helped build the models doing the work, he coined the term for this way of working, and he is now using it as his default rather than as a party trick. When someone with that resume says he has stopped writing code by hand for projects like this, the honest response is to figure out which of your own projects just moved into that category.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>andrej-karpathy</category><category>youtube</category><category>ai</category><category>artificial-intelligence</category><category>claude</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Cisco Gave One 27 Year Old Unlimited AI Tokens to Refactor 98% of Its Code</title><link>https://soursen.org/articles/cisco-gave-one-27-year-old-unlimited-ai-tokens-to-refactor-98-of-its/</link><guid isPermaLink="true">https://soursen.org/articles/cisco-gave-one-27-year-old-unlimited-ai-tokens-to-refactor-98-of-its/</guid><description>Cisco’s president gave a 27-year-old engineer unlimited AI tokens and 24 months to refactor 98% of the company’s legacy code, with 4 weeks to prove it can work at all. I think the headline number is a motivational target rather than an engineering estimate, and the more interesting story is what it says about who companies now trust with the keys.</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The number doing the rounds is 98%. As in, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Cisco&quot;&gt;Cisco&lt;/a&gt; wants to rewrite 98% of its legacy code with AI inside 24 months, and it has handed the job to a 27-year-old with a 2 person team and no cap on how many AI tokens he can burn. Said out loud, it sounds unhinged. A company with 40 years of shipping networking gear, running critical infrastructure for a large share of the internet, pointing an AI at almost all of its old code and telling one young engineer to go nuts.&lt;/p&gt;
&lt;p&gt;I dug into what was actually said, because the gap between the headline and the facts is the whole story. The 98% is a target somebody set, not a result anybody achieved, and the difference matters more than the number does.&lt;/p&gt;
&lt;h2 id=&quot;what-actually-happened&quot;&gt;What actually happened&lt;/h2&gt;
&lt;p&gt;Here is the verifiable core, reported first by Fortune and picked up widely. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://newsroom.cisco.com/c/r/newsroom/en/us/executives/jeetu-patel.html&quot;&gt;Jeetu Patel&lt;/a&gt;, Cisco’s president and chief product officer, met a young engineer named Padmanabhan Krishnamurthy, who goes by Paddy, in Cisco’s San Francisco office. Krishnamurthy pitched an idea, that a tiny team could refactor Cisco’s legacy codebase using AI. Patel liked it enough to back it hard. As &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.tomshardware.com/tech-industry/big-tech/cisco-gives-27-year-old-engineer-unlimited-ai-tokens-to-rewrite-98-percent-of-the-companys-legacy-code-in-24-months-president-plucks-former-startup-founder-telling-him-to-go-nuts&quot;&gt;Tom’s Hardware reported&lt;/a&gt;, Patel’s words were, I have given him unlimited tokens because he is going to refactor code, and you can go nuts.&lt;/p&gt;
&lt;p&gt;The target is 98% of Cisco’s legacy code, refactored within 24 months. The proof of concept window is 4 weeks. If those 4 weeks show the approach works, the longer effort gets the runway. If they do not, it does not.&lt;/p&gt;
&lt;p&gt;Krishnamurthy is 27. He joined Cisco in 2025 when it acquired &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://techcrunch.com/2025/11/17/cisco-acquires-translation-startup-ezdubs/&quot;&gt;EzDubs&lt;/a&gt;, a Y Combinator backed real time translation startup he co-founded in 2023, which Cisco folded into its Webex collaboration tools. Patel’s own description of where Krishnamurthy sat before this is the part I keep thinking about. He was 5 levels below his experience level, in Patel’s words buried away in the organization doing something unstrategic. So a founder who sold his company to Cisco was parked in a junior slot until the president pulled him out to run a moonshot.&lt;/p&gt;
&lt;h2 id=&quot;refactor-and-rewrite-are-different-jobs-and-the-swap-matters&quot;&gt;Refactor and rewrite are different jobs, and the swap matters&lt;/h2&gt;
&lt;p&gt;The viral version says rewrite. Patel said refactor. These are different jobs, and the difference is most of the argument.&lt;/p&gt;
&lt;p&gt;A rewrite means throwing out the old code and building the same features fresh, usually in a new language or structure. A refactor means restructuring the existing code so it is cleaner and easier to work on, without changing what it does. Refactoring is the more plausible target for AI, because the behavior is already defined and tested, and the model is reshaping code toward a known result rather than inventing one. A full rewrite of 98% of a 40 year old codebase in 24 months would be one of the most aggressive software projects ever attempted. A refactor of that scope is still enormous, but it lives in a different universe of risk.&lt;/p&gt;
&lt;p&gt;I think the slip from refactor to rewrite in the retelling is exactly how a reasonable internal bet becomes a viral story about recklessness. Both versions are ambitious. Only one of them is sane.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.Cw9wysLK_kvhBI.webp&quot; alt=&quot;The 2 readings of the same announcement. Refactor and rewrite describe very different jobs and very different risk.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The 2 readings of the same announcement. Refactor and rewrite describe very different jobs and very different risk. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.tomshardware.com/tech-industry/big-tech/cisco-gives-27-year-old-engineer-unlimited-ai-tokens-to-rewrite-98-percent-of-the-companys-legacy-code-in-24-months-president-plucks-former-startup-founder-telling-him-to-go-nuts&quot;&gt;Patel quotes via Tom’s Hardware, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;big-rewrites-have-a-graveyard-and-everyone-in-software-knows-it&quot;&gt;Big rewrites have a graveyard, and everyone in software knows it&lt;/h2&gt;
&lt;p&gt;The reason the rewrite framing scares engineers is that the industry has watched this movie before. The big bang rewrite, where you stop and rebuild a working system from scratch, is one of the most reliable ways to kill a software project. Netscape famously decided to rewrite its browser from the ground up in the late 1990s, spent 3 years on it, shipped nothing usable for most of that time, and handed Internet Explorer the market while it rebuilt. Joel Spolsky called that decision the single worst strategic mistake any software company can make, and the essay he wrote about it, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.joelonsoftware.com/2000/04/06/things-you-should-never-do-part-i/&quot;&gt;Things You Should Never Do&lt;/a&gt;, is still passed around 25 years later for a reason.&lt;/p&gt;
&lt;p&gt;The pattern repeats because old code encodes knowledge that nobody wrote down. Every strange branch and defensive check is a scar from a real failure. Throw the code away and you throw away the lessons with it, then spend years rediscovering them one outage at a time. This is exactly why the refactor framing matters so much for the Cisco bet. A refactor keeps the working system running and reshapes it in place, with the existing tests as a safety net. That is the approach with a track record of succeeding, and it is the one Patel’s actual words point to, whatever the headlines say.&lt;/p&gt;
&lt;h2 id=&quot;why-98-is-a-stretch-target-rather-than-a-measurement&quot;&gt;Why 98% is a stretch target rather than a measurement&lt;/h2&gt;
&lt;p&gt;Round numbers at the top of a range are almost never measurements. Nobody counted Cisco’s code, modeled the AI throughput, and arrived at 98.0% as the achievable figure. 98% is the number you pick when you want to say nearly all of it and make the team stretch. The honest reading is that Patel set an outcome so large it forces a different way of working, then gave one person the resources to find out how close you can actually get.&lt;/p&gt;
&lt;p&gt;That framing is common in ambitious engineering. You aim at a number you are not sure you can hit, because aiming at a safe number guarantees a safe result. The 4 week proof of concept is the tell. If this were a committed roadmap, you would not gate it on a month long trial. You gate it because you genuinely do not know yet whether a 2 person AI team can refactor at the scale claimed, and you want cheap evidence before you bet real time on it.&lt;/p&gt;
&lt;p&gt;So the correct way to read 98% in 24 months is as a question Cisco is paying to answer, not a promise it has made. The reports themselves say as much. None of them state how much code has actually changed, or whether the trial has succeeded.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.B3cVmVJm_Z1xh0rR.webp&quot; alt=&quot;How the bet is structured. A small cheap trial first, the big target only if it clears.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;818&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How the bet is structured. A small cheap trial first, the big target only if it clears. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.tomshardware.com/tech-industry/big-tech/cisco-gives-27-year-old-engineer-unlimited-ai-tokens-to-rewrite-98-percent-of-the-companys-legacy-code-in-24-months-president-plucks-former-startup-founder-telling-him-to-go-nuts&quot;&gt;Patel quotes via Tom’s Hardware, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-is-genuinely-new&quot;&gt;The part that is genuinely new&lt;/h2&gt;
&lt;p&gt;Strip the hype and something real is left. A president of a company Cisco’s size looked at a 27 year old who had been filed into a junior role, listened to a pitch that contradicted how his most senior people were thinking, and handed him an uncapped budget and a direct mandate. Patel said the most AI forward people on his own team were not thinking the way Krishnamurthy was.&lt;/p&gt;
&lt;p&gt;That is a bet on a person more than on a technology. The AI is the tool. The actual decision was to route resources and authority around the normal hierarchy to someone with no seniority and a contrarian idea. Patel has been open that this creates friction with long tenured staff, and his line on it is that tension is good, as long as it is tension between ideas rather than people. Whether that holds up when a 27 year old is reshaping code that veterans have maintained for 20 years is the real test, and it has nothing to do with tokens.&lt;/p&gt;
&lt;p&gt;Patel has also said he wants to find more early career staff who challenge established practices, which is the part of this I think generalizes beyond Cisco. The old logic was that you earn authority by putting in years, and the senior people decide what gets built. The bet Patel is making is that AI flattens that, because a junior person who is fluent with these tools can now attempt work that used to require a large experienced team, so the thing worth paying for shifts from accumulated seniority toward willingness to try the uncomfortable approach. If that bet pays off at Cisco, every large company watching will copy it, and the career ladder a lot of engineers have been climbing changes shape underneath them. That is a bigger story than one codebase.&lt;/p&gt;
&lt;h2 id=&quot;the-context-nobody-puts-next-to-it&quot;&gt;The context nobody puts next to it&lt;/h2&gt;
&lt;p&gt;This announcement did not happen in a vacuum. In May 2026 Cisco reported record quarterly revenue of $15.8 billion, up 12% from a year earlier, and in the same breath &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techradar.com/pro/we-are-making-clear-strategic-investments-cisco-cuts-4-000-jobs-even-as-ai-orders-surge&quot;&gt;announced it would cut fewer than 4,000 jobs&lt;/a&gt;, under 5% of its workforce, in the following quarter. On the same call it raised its full year AI orders target to $9 billion from $5 billion, and its AI revenue target to $4 billion from $3 billion. The money was not disappearing. It was moving.&lt;/p&gt;
&lt;p&gt;CEO Chuck Robbins was direct about the logic. He said the companies that win in the AI era will be the ones with the discipline to shift investment toward the strongest demand, which means making hard decisions about where to invest. Cisco framed the cuts as a reallocation toward AI infrastructure, silicon, optics, and security, and was careful to say it was not AI replacing workers. I take that framing seriously, and I also notice that the practical effect on a laid off engineer is identical either way.&lt;/p&gt;
&lt;p&gt;Put the facts side by side. Cisco shed thousands of roles while posting record numbers and tripling down on AI, then handed a single young engineer unlimited AI resources to do work that once needed large teams. That is the whole thesis of the current AI moment compressed into one company. Money and headcount are moving from many people doing the work to few people directing AI that does the work. The Cisco story is striking because it is so literal about it. One founder, 2 people, no token limit, against 98% of the code that a much larger organization wrote over decades.&lt;/p&gt;
&lt;p&gt;I am not going to pretend that is purely inspiring or purely grim. For Krishnamurthy it is the opportunity of a career. For the engineers whose decades of work is now the input to a refactoring agent, it reads differently. Both things are true at once.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CGMfCVH7_1sn0EP.webp&quot; alt=&quot;Cisco’s May 2026 restructuring and the code bet are the same strategy, money moving from headcount to AI.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;853&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Cisco’s May 2026 restructuring and the code bet are the same strategy, money moving from headcount to AI. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techradar.com/pro/we-are-making-clear-strategic-investments-cisco-cuts-4-000-jobs-even-as-ai-orders-surge&quot;&gt;Cisco Q3 FY2026 results and press coverage&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-risk-cisco-is-actually-taking&quot;&gt;The risk Cisco is actually taking&lt;/h2&gt;
&lt;p&gt;Legacy code at a company like Cisco is not a tidy app. It is decades of patches, undocumented assumptions, and quiet fixes that each solved a specific production fire. Much of the value in old code is invisible, a single odd looking line that exists because something broke at 3am in 2009 and this was the fix. An AI refactoring that code can easily clean away the weird line without knowing it was load bearing.&lt;/p&gt;
&lt;p&gt;This is not theoretical. In July 2025 an AI coding agent on the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.eweek.com/news/replit-ai-coding-assistant-failure/&quot;&gt;Replit platform&lt;/a&gt; deleted a live production database during an explicit code freeze, then generated fake data and misrepresented what it had done. The user had given it room to operate, which is the same posture as go nuts. Autonomous coding agents are powerful and genuinely useful, and they also do catastrophic things with total confidence, because they do not understand consequences, they predict text.&lt;/p&gt;
&lt;p&gt;Refactoring lowers this risk compared to a rewrite, since the tests that define correct behavior already exist and can catch regressions. It does not remove it. The failure mode at Cisco scale is not one deleted database. It is a subtle behavior change refactored into networking code that runs inside other companies’ infrastructure, discovered weeks later when something downstream breaks. The 4 week trial is partly about speed, and it should also be about proving the guardrails hold before anything touches code that matters.&lt;/p&gt;
&lt;h2 id=&quot;what-i-think-this-actually-is&quot;&gt;What I think this actually is&lt;/h2&gt;
&lt;p&gt;98% of the code rewritten by AI in 24 months is a headline. Refactoring a large legacy codebase with a small AI assisted team, gated on a 4 week proof, run by someone willing to challenge the senior engineers, is a real experiment, and a sane one to try.&lt;/p&gt;
&lt;p&gt;The number will almost certainly not land at 98%. That was never the point of a number like that. What Cisco is really testing is whether a tiny team with AI can do work that used to need a large one, and whether a company can hand that work to a 27 year old over the heads of people who have been there for decades. Those are the questions worth watching, and we will not have an answer from the headline. We will have it from whether the 4 weeks turn into 24 months, and from what actually ships.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>cisco</category><category>ai</category><category>artificial-intelligence</category><category>ai-agent</category><category>claude</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Chrome Spent 5 Years Telling Everyone Nobody Wanted JPEG XL. In 2026 It Shipped It Anyway</title><link>https://soursen.org/articles/chrome-spent-5-years-telling-everyone-nobody-wanted-jpeg-xl-in-2026-it/</link><guid isPermaLink="true">https://soursen.org/articles/chrome-spent-5-years-telling-everyone-nobody-wanted-jpeg-xl-in-2026-it/</guid><description>In 2022 Google pulled JPEG XL from Chrome, saying there was not enough interest from the ecosystem, while Meta, Adobe, Intel, Shopify, and The Guardian were on the record asking for it. Now Chrome 155 ships a JPEG XL decoder, rewritten in Rust, and the reasons the format was killed look very different in hindsight. I went through the whole timeline, the compression numbers, and who actually wanted this.</description><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;For about 5 years, the story of JPEG XL was a case study in how one company’s decision can stall a web standard that nearly everyone else wanted. Google added experimental support to Chrome, removed it while claiming the ecosystem was not interested, and spent years pointing people toward a competing format it had helped create. Then, in October 2026, Chrome 155 shipped JPEG XL decoding anyway, built on a new decoder written in Rust. The format that was supposedly unwanted is now in the browser that said nobody wanted it. I want to walk through how that happened, because the public reasons given at each step did not hold up.&lt;/p&gt;
&lt;h2 id=&quot;what-jpeg-xl-actually-is&quot;&gt;What JPEG XL actually is&lt;/h2&gt;
&lt;p&gt;JPEG XL is a modern image format standardized by the JPEG committee itself, with the core coding system finalized on 30 March 2022. By Chrome’s own description, it offers 30 to 50% better compression than ordinary JPEG, lossless compression, built-in HDR support, and progressive decoding, where a browser can show a low resolution version of an image first and sharpen it as more data arrives.&lt;/p&gt;
&lt;p&gt;The feature that matters most for the existing web is lossless JPEG transcoding. You can take the billions of ordinary JPEG files already out there and repack them as JPEG XL with no quality loss, shrinking them by around 20%, and convert them back to the exact original bytes later if needed. A format that can losslessly shrink the entire existing stock of web images by a fifth is not a small thing.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.D4iVSqGK_Zdl2Ny.webp&quot; alt=&quot;Figure 1. What JPEG XL offers over ordinary JPEG. Chrome cites 30 to 50% better compression, plus lossless recompression of existing JPEG files at around 20% smaller with a byte-exact way back, HDR, and progressive decoding.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1060&quot; height=&quot;575&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. What JPEG XL offers over ordinary JPEG. Chrome cites 30 to 50% better compression, plus lossless recompression of existing JPEG files at around 20% smaller with a byte-exact way back, HDR, and progressive decoding. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://developer.chrome.com/blog/jpeg-xl&quot;&gt;Author chart based on Chrome’s developer announcement&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;where-it-came-from&quot;&gt;Where it came from&lt;/h2&gt;
&lt;p&gt;JPEG XL is not a Google invention, which is part of why the story stings. It grew out of 2 separate projects that the JPEG committee merged. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://jpegxl.info/about/history.html&quot;&gt;Jon Sneyers&lt;/a&gt;, an engineer at the image company Cloudinary, built a lossless format called FLIF in 2015 and then a successor called FUIF, the Free Universal Image Format. Google, separately, built a format called PIK. When the JPEG committee put out a call for proposals in 2018, FUIF and PIK were 2 of the 7 submissions, and in January 2019 the committee chose to combine them. The 2 teams then spent about 2 years merging the designs into the single format.&lt;/p&gt;
&lt;p&gt;That history shows up in how the format works. JPEG XL has 2 modes under the hood. A VarDCT mode handles lossy photographic images using a block transform descended from the same math JPEG uses, but with variable block sizes from 2x2 up to 256x256, and it traces back to Google’s PIK. A modular mode handles lossless images, progressive loading, and extra channels like transparency, and it traces back to Sneyers’ FUIF. The compression gains come from combining these, which is why JPEG XL can cover everything from a tiny icon to a huge HDR photograph in a single format. The reference software that implements it is an open source library called libjxl.&lt;/p&gt;
&lt;h2 id=&quot;the-flag-then-the-reversal&quot;&gt;The flag, then the reversal&lt;/h2&gt;
&lt;p&gt;Chromium added JPEG XL support behind an experimental flag on 1 April 2021, around the same time Firefox did. It was not advertised, just a thing you could switch on to try. That lasted about 18 months. On 31 October 2022, a Google engineer posted the reasons Chrome would remove it, and the list is worth reading closely because each point was later undercut.&lt;/p&gt;
&lt;p&gt;The 4 stated reasons were that experimental flags should not remain indefinitely, that there was not enough interest from the entire ecosystem to keep experimenting with JPEG XL, that the format did not bring sufficient incremental benefits over existing formats to warrant enabling by default, and that removing the code in Chrome 110 would cut maintenance work and let Google focus on existing formats. Support was pulled from Chromium on 9 December 2022.&lt;/p&gt;
&lt;h2 id=&quot;the-ecosystem-that-supposedly-was-not-there&quot;&gt;The ecosystem that supposedly was not there&lt;/h2&gt;
&lt;p&gt;The second reason is the one that drew the most anger, because the interest was documented, in the same discussion thread, before the removal was announced. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/JPEG_XL&quot;&gt;Meta&lt;/a&gt; said on the record that Facebook was eagerly awaiting full JPEG XL support in Chrome. An engineer on Adobe’s Photoshop, Camera Raw, and Lightroom teams said JPEG XL was the best available codec for broad distribution of HDR still photos. Intel was on the thread. So were Shopify, Cloudinary, Flickr, SmugMug, the Krita Foundation, Serif, and The Guardian.&lt;/p&gt;
&lt;p&gt;That is a social network with billions of users, the company that makes the dominant professional image software, a chip maker that produces a large share of the world’s CPUs, a major e-commerce platform, and a national newspaper, all asking for the format, before Google said the ecosystem did not want it. Whatever ecosystem Google was describing, this was not a niche with no backers. The gap between the stated reason and the visible evidence is what turned a routine deprecation into a multi-year grievance.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.6nHSHcHe_Z2nL2j7.webp&quot; alt=&quot;Figure 2. A sample of the companies publicly backing JPEG XL in the Chromium thread before Google said there was not enough ecosystem interest. The list spans social media, creative software, chip makers, e-commerce, and news.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1220&quot; height=&quot;588&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. A sample of the companies publicly backing JPEG XL in the Chromium thread before Google said there was not enough ecosystem interest. The list spans social media, creative software, chip makers, e-commerce, and news. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/JPEG_XL&quot;&gt;Author chart based on the Chromium issue thread and reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-avif-question&quot;&gt;The AVIF question&lt;/h2&gt;
&lt;p&gt;A lot of the anger centered on a conflict of interest. The main rival to JPEG XL is AVIF, an image format derived from the AV1 video codec, and Google was heavily involved in creating AV1. Removing JPEG XL while steering developers toward AVIF looked, to a lot of people, like a company favoring its own format over a committee standard that outperformed it in important cases. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.fsf.org/blogs/community/google-our-browser-engine-overlords&quot;&gt;Free Software Foundation&lt;/a&gt; went further, with Greg Farough criticizing Google for holding a disturbing amount of control over the web.&lt;/p&gt;
&lt;p&gt;The performance picture was genuinely mixed, which muddied the argument. JPEG XL tends to win at higher quality settings, while AVIF often looks better at low quality, where it hides compression artifacts more effectively, and some of that gap comes down to the encoders, not the formats themselves. Jon Sneyers, one of JPEG XL’s creators, argued Google had misread its own test data. The honest summary is that the 2 formats traded wins depending on the test, which makes not bring sufficient incremental benefits a hard claim to defend when a major browser was simultaneously shipping the rival.&lt;/p&gt;
&lt;h2 id=&quot;the-rest-of-the-world-moved-without-google&quot;&gt;The rest of the world moved without Google&lt;/h2&gt;
&lt;p&gt;While Chrome sat out, the format kept spreading, and the spread reached well past other browsers. Apple added JPEG XL to Safari 17, which meant support across iPhones, iPads, and Macs in one move, a chunk of the market no smaller than the ecosystem Google claimed was missing. It went deeper than the browser too. The iPhone 16 Pro can capture its ProRAW photos as JPEG XL, and Samsung’s Expert RAW on Galaxy phones uses it as well, both built on the DNG 1.7 camera raw specification that added JPEG XL compression in 2023.&lt;/p&gt;
&lt;p&gt;The software support piled up in parallel. Adobe added JPEG XL to Lightroom and to Photoshop from version 26.8, and GIMP, Krita, Affinity, ImageMagick, and FFmpeg all handle it. Windows 11 supports it through a Microsoft image extension, and Linux desktops read it through GNOME, KDE, and COSMIC. In September 2025 the PDF Association named JPEG XL its preferred format for embedded HDR images. By the time Chrome reconsidered, the format was a normal option in cameras, phones, operating systems, and professional image tools. The only conspicuous hole in the map was the browser with the largest share.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BzKscXfA_Z2sLvyu.webp&quot; alt=&quot;Figure 3. JPEG XL adoption outside the browser while Chrome sat out. Apple ProRAW on the iPhone 16 Pro, Samsung Expert RAW, Windows 11, Adobe, GIMP, Krita, FFmpeg, and the PDF Association for HDR all supported it before Chrome shipped.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;985&quot; height=&quot;588&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. JPEG XL adoption outside the browser while Chrome sat out. Apple ProRAW on the iPhone 16 Pro, Samsung Expert RAW, Windows 11, Adobe, GIMP, Krita, FFmpeg, and the PDF Association for HDR all supported it before Chrome shipped. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/JPEG_XL&quot;&gt;Author chart based on JPEG XL adoption records&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Mozilla’s position shifted too. Firefox stayed neutral through 2023, unwilling to be the driving force but open to adopting JPEG XL if it gained traction. By 2024 Mozilla had focused its concern on the security of the original C++ reference decoder and signaled it would support a Rust implementation. That detail turned out to matter, because it previewed exactly the condition Chrome would later set.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.BgXc-ffW_Z49jUr.webp&quot; alt=&quot;Figure 4. The 5 year timeline. An experimental flag in 2021, removal in 2022, Safari shipping it in 2023, Chrome reversing its position in 2025, a Rust decoder merged, and Chrome 155 shipping it in 2026.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1074&quot; height=&quot;544&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The 5 year timeline. An experimental flag in 2021, removal in 2022, Safari shipping it in 2023, Chrome reversing its position in 2025, a Rust decoder merged, and Chrome 155 shipping it in 2026. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/JPEG_XL&quot;&gt;Author timeline based on Chromium, Safari, and Mozilla records&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-story-changed-from-no-one-wants-it-to-we-want-it-in-rust&quot;&gt;The story changed from no one wants it to we want it in Rust&lt;/h2&gt;
&lt;p&gt;After roughly 2 more years of silence, Chrome’s position moved. On 21 November 2025, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://groups.google.com/a/chromium.org/g/blink-dev/c/WjCKcBw219k&quot;&gt;Rick Byers&lt;/a&gt; of the Chrome team posted that since JPEG XL was last evaluated, Safari had shipped support, Firefox had updated its position to neutral and willing to adopt, and developer signals kept coming in through bug votes, interop proposals, and survey data, with JPEG XL also being added to PDF. Given those signals, Chrome would welcome contributions to integrate a performant, memory safe JPEG XL decoder, and with long term maintenance committed and the usual launch criteria met, it would ship it.&lt;/p&gt;
&lt;p&gt;The objection had quietly changed. It was no longer that the ecosystem did not care, because it plainly did. The new requirement was a decoder written in Rust, replacing the existing C++ one. That is a reasonable engineering ask on its own, image decoders are among the most attacked surfaces in a browser, processing untrusted binary data from the network, and memory bugs there are dangerous. The awkward part is that the organization asking for the Rust decoder, and saying it could not simply build one itself, is Google, whose employees lead the Chromium project. A company that size presenting itself as unable to resource the work read as disingenuous to many watching.&lt;/p&gt;
&lt;h2 id=&quot;how-it-finally-shipped&quot;&gt;How it finally shipped&lt;/h2&gt;
&lt;p&gt;Someone built the decoder. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://developer.chrome.com/blog/jpeg-xl&quot;&gt;jxl-rs&lt;/a&gt;, a JPEG XL decoder written in pure Rust, was merged into Chromium, and the flag became available in Chromium 145. To match the speed of the C++ version, the team leaned on recent Rust features, stabilizing the ability to run SIMD acceleration without unsafe code and building a SIMD abstraction layer called jxl_simd, modeled on the C++ Highway library. The result is a decoder that is roughly as fast as the memory unsafe alternative while removing the class of memory bugs that plague C++ decoders. Chrome’s team said fuzzing and AI assisted review found no memory safety issues in the implementation’s history.&lt;/p&gt;
&lt;p&gt;On 6 October 2026, Chrome engineers announced that Chrome 155 would ship JPEG XL decoding. The recommendation to developers is pragmatic, try both AVIF and JPEG XL and use whichever gives better results for a given image, with JPEG XL expected to help most for high fidelity or lossless photographic images and for progressive decoding. You serve it through a picture element with a fallback format, so browsers that do not support it still get a working image.&lt;/p&gt;
&lt;h2 id=&quot;what-this-actually-shows&quot;&gt;What this actually shows&lt;/h2&gt;
&lt;p&gt;The technology was never really the problem. The decoder could have been written in Rust years earlier, the compression gains were real the whole time, and the ecosystem that Google said was missing was sitting in the comment thread with its name attached. What the saga shows is how much power over a web standard sits with whoever controls the dominant browser, and how a stated reason can function as cover for a preference, in this case for a format the same company had a hand in.&lt;/p&gt;
&lt;p&gt;The part worth taking from it is that the pressure worked. Apple shipping Safari support, Mozilla moving to neutral, and years of developers refusing to let the issue die left Chrome with a position that was hard to hold. Plenty of formats die quietly when Google declines to support them. This one did not, because the people who wanted it kept showing up, and eventually the browser that said no had to change its answer.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>chrome</category><category>google</category><category>web-development</category><category>web3</category><category>browsers</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>220,000 Developers Starred a Text File.</title><link>https://soursen.org/articles/220-000-developers-starred-a-text-file/</link><guid isPermaLink="true">https://soursen.org/articles/220-000-developers-starred-a-text-file/</guid><description>Let me tell you about the most-starred AI repository of 2026.</description><pubDate>Sun, 28 Jun 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Let me tell you about the most-starred AI repository of 2026.&lt;/p&gt;
&lt;p&gt;It has no code. No dependencies. No installation script. No framework to learn. It is a single plain text file, 70 lines long, that you drop into the root of your project. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/forrestchang/andrej-karpathy-skills&quot;&gt;The repository forrestchang/andrej-karpathy-skills&lt;/a&gt; has accumulated over 220,000 combined GitHub stars across two accounts approximately 91,200 on Forrest Chang’s personal account and over 132,000 on the organisation mirror at multica-ai/andrej-karpathy-skills since it was created on January 27, 2026. It appeared on GitHub Trending for weeks on end, with the organisation mirror reaching the number one position on April 8, 2026. It has been installed by developers across every major language, framework, and industry.&lt;/p&gt;
&lt;p&gt;The file is called CLAUDE.md. It contains 4 behavioral rules for AI coding agents, when I first read those rules, my immediate reaction was that these are obvious. Every senior engineer would tell a junior these things on day one. Why does a text file stating obvious things become the most viral AI repository of the year?&lt;/p&gt;
&lt;p&gt;The answer to that question is more interesting than the file itself.&lt;/p&gt;
&lt;h2 id=&quot;where-this-came-from-and-why-it-spread-the-way-it-did&quot;&gt;Where This Came From and Why It Spread the Way It Did&lt;/h2&gt;
&lt;p&gt;On January 26, 2026, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://x.com/karpathy/status/2015883857489522876&quot;&gt;Andrej Karpathy posted on X&lt;/a&gt; about what he called the biggest change to his basic coding workflow in two decades of programming. Karpathy is not a casual observer of this space. He is a co-founder of OpenAI, the former Director of AI at Tesla, a Stanford PhD, and the person who coined the term vibe coding in early 2025. When he says something changed his workflow, people pay attention.&lt;/p&gt;
&lt;p&gt;What he described was a shift from 80% manual coding with autocomplete to 80% agent-driven development in roughly 2 months. He had made the same transition that millions of developers were making around the same time, following what he described as a phase shift in software engineering that happened around December 2025, when agentic coding tools crossed a threshold of coherence that made them genuinely useful for production work rather than just demos.&lt;/p&gt;
&lt;p&gt;But the post was not celebratory. It was a precise inventory of failure modes. Karpathy wrote: &lt;em&gt;“They don’t manage their confusion, don’t seek clarifications, don’t surface inconsistencies, don’t present tradeoffs, don’t push back when they should.”&lt;/em&gt; He identified 4 specific recurring patterns of failure that he kept encountering. Silent wrong assumptions. Overcomplicated solutions with unnecessary abstractions. Unauthorised edits to code adjacent to the requested task. Vague instructions being executed without defining what success would look like.&lt;/p&gt;
&lt;p&gt;The post gathered significant attention across the developer community. The following day, developer Forrest Chang did something practical with it. He took Karpathy’s observations and encoded them into a CLAUDE.md file, 4 named principles each addressing one failure mode Karpathy had identified. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techtimes.com/articles/316798/20260518/karpathy-inspired-claudemd-passes-220000-combined-github-stars-four-rules-that-stop-ai-breaking.htm&quot;&gt;The repo crossed 43,000 stars in its first week&lt;/a&gt;, reached 110,000 by early May on the personal account alone, and passed 220,000 combined stars with the organisation mirror by mid-May 2026.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://survey.stackoverflow.co/2025/technology&quot;&gt;Stack Overflow’s 2025 developer survey&lt;/a&gt; found that 84% of developers now use AI coding tools, but trust in AI output had fallen from 40% to 29% year over year. That gap between usage and trust is the precise tension the CLAUDE.md file addresses. People kept using the tools because the productivity benefit was real. People stopped trusting the output because the failure modes were predictable and consistent. Karpathy named those failure modes precisely. Chang turned the names into a configuration file. The community recognised both in the same instant.&lt;/p&gt;
&lt;h2 id=&quot;the-4-rules-and-why-each-one-exists&quot;&gt;The 4 Rules and Why Each One Exists&lt;/h2&gt;
&lt;p&gt;Before going into each rule, it helps to understand what a CLAUDE.md file actually is. Claude Code, Anthropic’s coding agent, reads any file named CLAUDE.md in the project root at the start of every session. It treats the contents as persistent context for that project. The file gets committed to version control, so every developer on the team gets the same behavioral guardrails automatically. Zero configuration beyond creating the file.&lt;/p&gt;
&lt;p&gt;The karpathy-skills CLAUDE.md does not contain project-specific instructions about naming conventions or testing frameworks. It contains behavioral principles generic enough to drop into any codebase and immediately change how the agent approaches tasks.&lt;/p&gt;
&lt;h2 id=&quot;rule-one-think-before-coding&quot;&gt;Rule One: Think Before Coding&lt;/h2&gt;
&lt;p&gt;This rule addresses what researchers call the silent assumption problem. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://byteiota.com/karpathy-claude-md-ai-coding-pitfalls-accuracy-2/&quot;&gt;A 2025 academic study cited by ByteIota&lt;/a&gt; found that 94% of LLM compilation errors are type-check failures, where the model assumed a type rather than asking for clarification. The pattern is consistent across every kind of ambiguity, not just types. You ask the agent to add validation. It picks a validation level without asking which one you need. You ask it to refactor a function. It picks an approach without checking whether you have constraints that would rule it out.&lt;/p&gt;
&lt;p&gt;The rule requires the agent to state its assumptions explicitly before writing any code, present multiple interpretations when a request is ambiguous, and ask rather than guess when something is genuinely unclear. One good clarifying question at the start of a task eliminates five rounds of revision at the end. The principle sounds obvious because it is. The model does not follow it by default. That gap is the entire problem.&lt;/p&gt;
&lt;h2 id=&quot;rule-two-simplicity-first&quot;&gt;Rule Two: Simplicity First&lt;/h2&gt;
&lt;p&gt;Karpathy described this failure directly in his post: &lt;em&gt;“They really like to overcomplicate code and APIs, bloat abstractions, don’t clean up dead code… implement a bloated construction over 1,000 lines when 100 would do.”&lt;/em&gt; The rule requires the agent to write the minimum code that solves the problem, avoid speculative abstractions, and treat over-engineering as a failure rather than thoroughness. The test: would a senior engineer call this overcomplicated? If yes, rewrite it.&lt;/p&gt;
&lt;h2 id=&quot;rule-three-surgical-changes&quot;&gt;Rule Three: Surgical Changes&lt;/h2&gt;
&lt;p&gt;This is the rule that addresses what Karpathy called collateral damage that the agent editing code it was never asked to touch. I have experienced this in specific enough ways that I can describe it precisely. You ask an agent to fix a date parsing bug. It fixes the bug. It also reformats the entire file, switches unrelated functions from list comprehensions to loops, adds type hints to functions in a different module, and renames a variable it thought was unclear. None of that was requested. All of it now requires review.&lt;/p&gt;
&lt;p&gt;Karpathy described this pattern exactly: &lt;em&gt;“They still sometimes change/remove comments and code they don’t sufficiently understand as side effects, even if orthogonal to the task.”&lt;/em&gt; The rule requires the agent to touch only what the task requires, leave neighbouring code exactly as found, and trace every changed line back to the original request. If a line changed and cannot be traced to the task, it should not have changed.&lt;/p&gt;
&lt;h2 id=&quot;rule-four-goal-driven-execution&quot;&gt;Rule Four: Goal-Driven Execution&lt;/h2&gt;
&lt;p&gt;This rule addresses the absence of verifiable success criteria. Vague instructions like add validation or fix the bug give the agent no way to know when it has finished, which produces either premature stops or endless iteration. As the CLAUDE.md &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://medium.com/@tentenco/the-claude-md-guide-how-one-file-turns-a-stateless-ai-into-your-long-term-coding-partner-b8806683df7f&quot;&gt;guide on Medium describes it&lt;/a&gt;, clear success criteria let the model iterate autonomously. Vague criteria require intervention at every step.&lt;/p&gt;
&lt;p&gt;The rule requires the agent to turn vague instructions into verifiable targets before writing a line. Add validation becomes write tests for invalid inputs, then make them pass. Fix the bug becomes write a test that reproduces the issue, then make the test pass. The shift from imperative to goal-based is not stylistic. It changes the model’s ability to evaluate its own output.&lt;/p&gt;
&lt;h2 id=&quot;what-the-star-count-is-actually-measuring&quot;&gt;What the Star Count Is Actually Measuring&lt;/h2&gt;
&lt;p&gt;220,000 combined stars on a plain text file is not a vote on the file. It is a vote on the problem. Every developer who read Karpathy’s post recognised the failures he described from their own sessions. The file went viral because the bar for adoption was that can I copy-paste this into my repo right now. The answer was yes. So hundreds of thousands of people did.&lt;/p&gt;
&lt;p&gt;The context matters. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://byteiota.com/karpathy-claude-md-ai-coding-pitfalls-accuracy-2/&quot;&gt;ByteIota’s analysis of AI coding adoption&lt;/a&gt; found that 68% of developers now use AI to generate code, with 1.1 million public repositories importing an LLM SDK, up 178% year over year. Nearly 80% of new developers use Copilot within their first week. Karpathy himself noted that 2026 was on track to become what he called the slopacolypse: a flood of low-quality AI-generated code produced by agents operating without behavioral guardrails.&lt;/p&gt;
&lt;p&gt;TypeScript becoming GitHub’s number one language in 2026, with 66% growth driven by 2.6 million monthly contributors, is partly a response to this. When AI writes more of your code, type safety matters more, because the agent makes type errors consistently, and a type system catches them before they ship. The four rules in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://claude.md/&quot;&gt;CLAUDE.md&lt;/a&gt; are addressing the same problem at the behavioral layer rather than the language layer.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;The most viral AI tool of the month was not a new model. It was a document telling an existing model to behave like a thoughtful engineer rather than a confident one.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;the-honest-caveats-worth-stating-clearly&quot;&gt;The Honest Caveats Worth Stating Clearly&lt;/h2&gt;
&lt;p&gt;The file is behavioral context, not a contract. Claude Code reads CLAUDE.md and treats it as instruction. It does not always follow it. The principles improve the distribution of behavior. They do not guarantee specific behavior on any given task. Reddit threads pointing this out are correct.&lt;/p&gt;
&lt;p&gt;The accuracy improvement numbers that circulate online come from &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://byteiota.com/karpathy-claude-md-ai-coding-pitfalls-accuracy-2/&quot;&gt;specific skill evaluations documented by ByteIota&lt;/a&gt; covering particular domains. Those are not general coding accuracy figures and should not be treated as such. The directional claim, that the file improves output quality meaningfully, is supported by widespread developer reports. The specific numbers are context-dependent.&lt;/p&gt;
&lt;p&gt;The principles are not Claude-specific. The same 4 rules, with the filename changed to match the tool, work as Cursor rules or as a system prompt prefix for any coding agent. The viral repo ships both versions. The interesting thing is not the specific tool integration. It is the principles themselves, which describe how any coding agent should behave when given real work in a real codebase.&lt;/p&gt;
&lt;p&gt;Karpathy did not write the file. He posted the observations on January 26. Forrest Chang turned them into a repo the following day. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techtimes.com/articles/316798/20260518/karpathy-inspired-claudemd-passes-220000-combined-github-stars-four-rules-that-stop-ai-breaking.htm&quot;&gt;A Snyk audit of the agent skills ecosystem published in February 2026&lt;/a&gt; found that 13% of agent skills packages contained critical security flaws. Security researchers at Adversa AI and LayerX additionally documented that malicious CLAUDE.md files planted in cloned repositories can instruct Claude Code to exfiltrate SSH keys and API credentials. The practical implication is simple that install the file only from the official repositories, specifically forrestchang/andrej-karpathy-skills or multica-ai/andrej-karpathy-skills. The file is 70 lines of plain text, human-readable end to end, and takes under a minute to inspect before installing.&lt;/p&gt;
&lt;p&gt;The repo is a starting point, not a complete configuration. The four rules work as a behavioral baseline. They do not replace project-specific instructions about naming conventions, testing frameworks, or architecture decisions. The right approach is to append a behavioral guidelines section to whatever CLAUDE.md you already have rather than replacing it with the Karpathy file.&lt;/p&gt;
&lt;h2 id=&quot;the-deeper-point-nobody-is-talking-about&quot;&gt;The Deeper Point Nobody Is Talking About&lt;/h2&gt;
&lt;p&gt;The energy in AI tooling right now goes into more layers. More model capability. More context window. More integrations. More plugins. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://miraflow.ai/blog/karpathy-claude-md-100k-github-stars-ai-coding-2026&quot;&gt;At Anthropic itself&lt;/a&gt;, the majority of code is now written by Claude Code, with engineers focusing on architecture, product thinking, and continuous orchestration rather than implementation. The trajectory of the industry is toward agents that do more, autonomously, with less human intervention.&lt;/p&gt;
&lt;p&gt;The CLAUDE.md phenomenon runs in the opposite direction. It is not adding capability. It is adding constraint. It is saying: before the agent does anything, it should think. Before it writes, it should ask. Before it changes, it should verify the scope. Before it finishes, it should confirm it has met a verifiable goal.&lt;/p&gt;
&lt;p&gt;These are not novel ideas. They are what thoughtful human engineers do automatically, after years of experience teaching them what the consequences of not doing those things look like. The gap between a model that has processed more code than any human has ever read and a model that behaves like an experienced engineer is precisely this: the experienced engineer knows when to stop, when to ask, when the solution is too complicated, when the scope has drifted. The model does not know those things by default.&lt;/p&gt;
&lt;p&gt;The best coding agent improvement of 2026 so far is a markdown file. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://medium.com/@aitrends24/i-installed-the-144k-star-karpathy-claude-md-file-heres-what-changed-d3dd31b8858d&quot;&gt;One developer who installed it&lt;/a&gt; described the setup time as four minutes and said the behavioral change was noticeable from the first task. Another described months of correcting the same session-to-session mistakes eliminated overnight by a 70-line file.&lt;/p&gt;
&lt;p&gt;The lesson here is not that configuration beats capability. It is that capability without behavioral structure produces the specific failure modes Karpathy documented. The model is capable enough to do good work. The default behavior does not consistently produce good work. A plain text file closing that gap is not a sign that the model is limited. It is a sign that the model is powerful enough that the difference between directed and undirected is the entire gap between good output and frustrating output.&lt;/p&gt;
&lt;p&gt;If you use Claude Code for any real work, the file is worth installing. Not because it is magic. Because the four problems it solves are the four problems you have already encountered, and naming them precisely is more than half of fixing them.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>github</category><category>claude</category><category>ai</category><category>artificial-intelligence</category><category>markdown</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>LinkedIn Is Broken. And You Already Know It.</title><link>https://soursen.org/articles/linkedin-is-broken-and-you-already-know-it/</link><guid isPermaLink="true">https://soursen.org/articles/linkedin-is-broken-and-you-already-know-it/</guid><description>A honest, uncomfortable look at what the platform has become and what to do instead.</description><pubDate>Mon, 01 Jun 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Let me paint you a picture. You’ve spent 3 evenings this week rewriting your resume. You’ve tailored your cover letter for each application. You’ve applied to 52 jobs on LinkedIn over the last six weeks. You’ve heard back from 4. Two of those were automated rejections. One was a recruiter asking if you were available for a role that had nothing to do with what you applied for. And the last one? That listing was posted in September. The role was filled in October. It’s January.&lt;/p&gt;
&lt;p&gt;Sound familiar? You’re not alone and more importantly, it’s not you.&lt;/p&gt;
&lt;p&gt;LinkedIn, the platform that promised to be the world’s most useful professional network, has turned into something else entirely. It’s part job board with phantom listings, part Instagram for people in blazers, part scam vector, and part motivation-poster factory. And because so many of us depend on it for jobs, visibility, professional credibility we keep using it even when it’s clearly not working.&lt;/p&gt;
&lt;p&gt;So let’s talk about it properly. What’s broken, why it’s broken, and most importantly what you should actually be doing instead.&lt;/p&gt;
&lt;h2 id=&quot;the-ghost-job-problem-is-worse-than-you-think&quot;&gt;The Ghost Job Problem Is Worse Than You Think&lt;/h2&gt;
&lt;p&gt;Start here, because this one affects the most people and causes the most quiet, grinding harm.&lt;/p&gt;
&lt;p&gt;Ghost jobs listings that are already filled, never existed, or are being kept live on purpose have become a defining feature of the modern job board experience and LinkedIn is drowning in them. A 2024 survey by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.resumebuilder.com/3-in-10-companies-currently-have-fake-job-posting-listed/&quot;&gt;ResumeBuilder.com&lt;/a&gt; which surveyed 1,641 hiring managers found that &lt;strong&gt;40% of companies admitted to posting a fake job listing in the previous year.&lt;/strong&gt; 3 in 10 companies still had fake listings live at the time of the survey. The reasons they gave were remarkable in their casualness: ‘collecting CVs for future openings,’ ‘making employees feel replaceable,’ ‘giving the impression the company is growing.’ 1 in 10 companies admitted they had no intention of filling the role they were actively advertising.&lt;/p&gt;
&lt;p&gt;It gets worse. A &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.fastcompany.com/91425252/recruiters-dish-on-ghost-jobs-why-companies-post-them-and-how-to-outsmart-them&quot;&gt;2025 Clarify Capital study&lt;/a&gt; found that nearly 1 in 3 employers admit to posting jobs with no intention to hire.Meanwhile, a 2024 analysis of &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.yahoo.com/news/employees-navigate-increasingly-soul-crushing-090000483.html&quot;&gt;Greenhouse’s platform data found&lt;/a&gt; that &lt;strong&gt;18–22% of all online job ads are fake or unfilled, with 70% of companies posting at least one ghost job in Q2 2024 alone.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;A separate 2024 analysis of almost 270,000 Glassdoor reviews by a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://vidcruiter.com/resources/fake-and-ghost-jobs/&quot;&gt;City University of New York PhD student&lt;/a&gt; estimated that up to 21% of job ads may be ghost jobs. And a 2025 analysis of LinkedIn job posts specifically suggested &lt;strong&gt;27.4% of US ads were ghost jobs&lt;/strong&gt; over one in four.&lt;/p&gt;
&lt;p&gt;What this means in practice? Every time you agonise over a tailored cover letter, research the company, customise your portfolio and press Apply there’s a statistically meaningful chance the role doesn’t exist. You are pouring real emotional labour and real time into a void.&lt;/p&gt;
&lt;p&gt;And LinkedIn profits from this. More listings mean more perceived opportunity. More perceived opportunity means more users. More users mean more ad revenue. There is no real business incentive to fix it aggressively. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.fortune.com/2024/06/26/bosses-lying-to-burned-out-employees-posting-fake-jobs&quot;&gt;As one recruiter put it to Fortune&lt;/a&gt; , posting fake jobs can cause ‘irreparable damage’ to employer brand yet 7 in 10 hiring managers say they believe the practice is morally acceptable.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;This isn’t a bug. It’s a business model feature that happens to cost you personally.&lt;/em&gt;&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;&lt;strong&gt;The ghost job numbers, at a glance: &lt;br/&gt;&lt;/strong&gt;&lt;/em&gt;
&lt;em&gt;→ 40% of companies posted a fake job listing in 2024 (ResumeBuilder, 1,641 hiring managers surveyed) &lt;br/&gt;
→ 27.4% of LinkedIn US job ads estimated to be ghost jobs (&lt;/em&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://resumeup.ai/&quot;&gt;&lt;em&gt;ResumeUp.AI&lt;/em&gt;&lt;/a&gt;&lt;em&gt;, 2025) &lt;br/&gt;
→ 18–22% of Greenhouse platform listings classified as fake or unfilled (Greenhouse internal data, Q2 2024)&lt;br/&gt;
→ 85% of companies posting fake jobs went so far as to interview candidates for non-existent roles &lt;br/&gt;
→ 3 states (New Jersey, California, Kentucky) have introduced legislation requiring removal of inactive listings&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;its-turned-into-instagram-and-not-the-good-parts&quot;&gt;It’s Turned Into Instagram. And Not the Good Parts.&lt;/h2&gt;
&lt;p&gt;Remember the old LinkedIn? The mildly boring but genuinely useful place where you updated your job title and occasionally endorsed someone for ‘Microsoft Excel’?&lt;/p&gt;
&lt;p&gt;That LinkedIn is gone. What replaced it is something genuinely strange.&lt;/p&gt;
&lt;p&gt;LinkedIn’s engagement algorithm rewards emotional, personal content above almost everything else. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://mercermackay.com/thinking/blog/a-leaders-guide-to-the-linkedin-algorithm-what-the-data-says/&quot;&gt;A 2025 analysis of 1.8 million LinkedIn posts&lt;/a&gt; by researcher Richard van der Blom found that organic reach has dropped by nearly 50%, and that visibility is now heavily concentrated among a smaller group of ‘top creators’ people who post frequently and generate emotional engagement.&lt;/p&gt;
&lt;p&gt;So professionals many of them smart, capable people who understand incentives have learned to reverse-engineer that. The result is a content genre that is uniquely, almost artistically insufferable.&lt;/p&gt;
&lt;p&gt;You’ve seen it. The founder who posts a photo of their Range Rover ‘not to flex, but to show what consistent effort looks like.’ The CEO who shares a story about crying on a flight and turns it into a 900-word post about ‘authentic leadership’ that ends with ‘follow for more real talk.’ The person who ‘almost quit last Tuesday’ but then saw something inspiring at the gym and decided to keep going and would you like to know their 5 steps to resilience? They have a course. Of course they have a course.&lt;/p&gt;
&lt;p&gt;This genre has a name, &lt;strong&gt;hustle porn.&lt;/strong&gt; It’s the LinkedIn-specific fusion of performative ambition, manufactured vulnerability, and vague inspiration optimised for reactions over insight. The platform’s algorithm has trained people to perform, and now the performers are everywhere.&lt;/p&gt;
&lt;p&gt;To be fair, LinkedIn’s 2026 algorithm updates are attempting to correct this &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.digitalapplied.com/blog/linkedin-algorithm-2026-engagement-strategy-guide&quot;&gt;Digital Applied&lt;/a&gt; reports that the platform is now actively suppressing ‘engagement bait’ and rewarding content depth over emotional manipulation. But the years of conditioning have left a feed that still feels far more like a motivational poster factory than a professional network.&lt;/p&gt;
&lt;h2 id=&quot;the-fake-profile-problem-is-massive-and-underreported&quot;&gt;The Fake Profile Problem Is Massive and Underreported&lt;/h2&gt;
&lt;p&gt;LinkedIn claims over one billion members. But how many of those are real, active, legitimate professionals?&lt;/p&gt;
&lt;p&gt;The honest answer? significantly fewer than a billion. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://besedo.com/blog/linkedin-fake-accounts/&quot;&gt;Besedo’s analysis of LinkedIn’s transparency data&lt;/a&gt;, LinkedIn blocked or removed &lt;strong&gt;more than 121 million fake accounts in 2023 alone.&lt;/strong&gt; In 2024, that number climbed further LinkedIn’s own transparency report confirmed &lt;strong&gt;100 million fake accounts removed in 2024, 80.6 million at registration stage, 19.7 million intercepted proactively, and 265,700 after user reports.&lt;/strong&gt; LinkedIn’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://about.linkedin.com/transparency/community-report&quot;&gt;official Community Report&lt;/a&gt; confirms 99.7% are caught proactively but that still leaves hundreds of thousands that aren’t.&lt;/p&gt;
&lt;p&gt;These aren’t just spam bots. The sophisticated fakes are genuinely hard to spot. AI-generated profile photos with that eerily smooth skin and slightly blurred background. Fabricated work histories pulling real company names. Profiles with hundreds of connections and endorsed skills that were gamed through fake account networks.&lt;/p&gt;
&lt;p&gt;What are these fake profiles used for?&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Recruiting scams that collect CVs and personal data, then disappear&lt;/li&gt;
&lt;li&gt;‘Pig butchering’ investment fraud, a fake persona spends weeks building a genuine-feeling relationship before introducing a ‘great opportunity’ in crypto or forex&lt;/li&gt;
&lt;li&gt;Competitor intelligence, fake recruiter profiles used to extract information from rival company employees&lt;/li&gt;
&lt;li&gt;Engagement farming, fake accounts that inflate post metrics and signal false popularity&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The FBI has formally identified LinkedIn as a significant vector for financial fraud. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.cnbc.com/2022/06/17/fbi-says-fraud-on-linkedin-a-significant-threat-to-platform-and-consumers.html&quot;&gt;In an exclusive interview with CNBC&lt;/a&gt;, FBI Special Agent Sean Ragan called crypto investment scams on LinkedIn &lt;strong&gt;‘a significant threat’ to both the platform and consumers.&lt;/strong&gt; The FBI’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.fbi.gov/how-we-can-help-you/victim-services/national-crimes-and-victim-resources/operation-level-up&quot;&gt;Operation Level Up&lt;/a&gt;, launched in 2024, was specifically aimed at intercepting victims of ‘pig butchering’ cryptocurrency scams that typically begin on professional platforms like LinkedIn.&lt;/p&gt;
&lt;p&gt;LinkedIn is trusted precisely because it’s professional. Your real name. Your employment history. Your credentials. The social proof is built into the design. Which is exactly why it’s such an attractive hunting ground for people who exploit trust.&lt;/p&gt;
&lt;h2 id=&quot;the-open-to-work-banner-well-intentioned-complicated-in-practice&quot;&gt;The ‘Open to Work’ Banner: Well-Intentioned, Complicated in Practice&lt;/h2&gt;
&lt;p&gt;When LinkedIn introduced the green ‘Open to Work’ banner, the intent was good, give job seekers a visible way to signal availability so recruiters could find them faster.&lt;/p&gt;
&lt;p&gt;The reality has been messier.&lt;/p&gt;
&lt;p&gt;Multiple recruiters and hiring managers many writing openly on LinkedIn itself, with no apparent awareness of the irony have admitted that they treat the banner as a &lt;strong&gt;negative signal.&lt;/strong&gt; The logic, as toxic as it is: if this person were truly talented and in-demand, someone would have already hired them. The banner signals desperation. Desperation is unattractive.&lt;/p&gt;
&lt;p&gt;LinkedIn’s own response was telling, they introduced a ‘private’ mode where the signal is only visible to recruiters, not your full network. A useful feature. But the fact that they needed to build it reveals something ugly about the platform’s culture.&lt;/p&gt;
&lt;p&gt;It has become a network where you must perform continuous professional success to maintain visibility. You can’t look uncertain. You can’t look like you’re struggling. You can’t be visibly between things. And that means the people who need LinkedIn most the ones actively looking are systematically disadvantaged for showing that need.&lt;/p&gt;
&lt;h2 id=&quot;the-influencer-economy-has-arrived-and-it-brought-all-its-distortions&quot;&gt;The Influencer Economy Has Arrived, and It Brought All Its Distortions&lt;/h2&gt;
&lt;p&gt;LinkedIn now has full-blown influencers. Not in a small, niche way, in a millions-of-followers, selling-courses, building-personal-empires way.&lt;/p&gt;
&lt;p&gt;There are accounts posting exclusively about ‘the habits of successful people’ and ‘what I learned from my first million.’ They sell coaching programmes, group masterminds, and PDF guides to ‘authentic networking.’ This wouldn’t be so bad if the platform didn’t actively amplify them over the mid-level engineer sharing a genuinely useful technical breakdown.&lt;/p&gt;
&lt;p&gt;The reach mechanics are the problem.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://mercermackay.com/thinking/blog/a-leaders-guide-to-the-linkedin-algorithm-what-the-data-says/&quot;&gt;Research by Richard van der Blom analysing 1.8 million posts&lt;/a&gt; found that ‘Top Creator’ visibility on LinkedIn climbed from 15% in 2022 to 31% in 2025, while visibility for other creators collapsed from 57% to 28%. Posting frequency and emotional resonance correlate with reach more than expertise or usefulness. The people who shout loudest about success are, in many cases, selling &lt;strong&gt;the idea of success&lt;/strong&gt; as their primary product.&lt;/p&gt;
&lt;p&gt;Meanwhile, the people who actually need LinkedIn, the ones looking for work, trying to recruit, hoping to make a real professional connection are sifting through an increasingly cluttered feed to find signal in the noise.&lt;/p&gt;
&lt;h2 id=&quot;linkedin-is-also-quietly-becoming-a-social-status-marker&quot;&gt;LinkedIn Is Also Quietly Becoming a Social Status Marker&lt;/h2&gt;
&lt;p&gt;Here’s something that doesn’t get discussed enough in Western tech circles, LinkedIn’s role has quietly expanded beyond professional utility in several cultures and contexts.&lt;/p&gt;
&lt;p&gt;In India where LinkedIn has over 100 million users and is growing faster than almost anywhere else the platform has developed a social dimension that goes well beyond job searching. It’s not uncommon for families to check a potential match’s LinkedIn profile as part of an informal background assessment. Job title, company name, degree institution, visible career progression, these have become markers of social credibility that extend into personal life.&lt;/p&gt;
&lt;p&gt;It’s not just India. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://restofworld.org/2025/linkedin-job-scams/&quot;&gt;Rest of World’s reporting on LinkedIn scam patterns globally&lt;/a&gt;, the platform’s perceived legitimacy and professional trust signals are precisely what make it so valuable and so exploitable across vastly different cultural contexts.&lt;/p&gt;
&lt;p&gt;Your LinkedIn profile has become a public reputation dashboard. Companies vet candidates before interviews. Clients research consultants before signing. Investors scan founders before taking meetings. And for many people globally, the pressure to perform on LinkedIn isn’t just professional anxiety it’s genuinely tied to life outcomes in a broader sense. Which makes the platform’s dysfunction feel less like an annoying product issue and more like a systemic problem.&lt;/p&gt;
&lt;h2 id=&quot;to-be-fair-what-linkedin-is-still-actually-good-for&quot;&gt;To Be Fair: What LinkedIn Is Still Actually Good For&lt;/h2&gt;
&lt;p&gt;I’m not here to tell you to delete your account. LinkedIn is still useful just not in the way most people use it. Here’s where it genuinely delivers:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Warm introductions.&lt;/strong&gt; If someone you know can connect you to someone at a company you’re targeting, LinkedIn is still the easiest way to make that happen.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Passive discovery.&lt;/strong&gt; A well-optimised profile makes it possible for recruiters doing active searches to find you especially with specific, in-demand skills.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Company research.&lt;/strong&gt; Before an interview, before a pitch, before a client meeting understanding structure, finding mutual connections, spotting recent moves.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Niche professional communities.&lt;/strong&gt; In some industries, substantive conversations still happen in comments and groups. You have to work to find them, but they exist.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Credibility signalling.&lt;/strong&gt; A complete, professional LinkedIn profile is simply expected in most industries. Not having one raises questions.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The key mindset shift is treat LinkedIn as a tool with specific jobs to do, not a platform to inhabit. Go in with a purpose, extract the value, leave. Don’t scroll. Don’t compare. Don’t interpret silence as a verdict on your worth.&lt;/p&gt;
&lt;h2 id=&quot;if-not-linkedin-then-what-heres-where-to-actually-find-good-opportunities&quot;&gt;If Not LinkedIn, Then What? Here’s Where to Actually Find Good Opportunities&lt;/h2&gt;
&lt;p&gt;This is the part most articles skip. They’ll tell you LinkedIn is broken and leave you staring at the wall. So let’s get specific.&lt;/p&gt;
&lt;p&gt;The honest truth is that the best opportunities the ones that result in real offers, at real companies rarely come from cold applications on a job board. They come from being in the right place, with the right visibility, before the job is even posted. Here’s how to build that.&lt;/p&gt;
&lt;h2 id=&quot;1-x-twitter---where-the-jobs-get-posted-before-theyre-posted&quot;&gt;1. X (Twitter) - Where the Jobs Get Posted Before They’re Posted&lt;/h2&gt;
&lt;p&gt;X has become one of the most effective job sourcing platforms for tech, design, marketing, startups, and creative roles specifically because the culture rewards transparency and directness in a way LinkedIn doesn’t. Founders tweet about what they’re building. Hiring managers share what they’re looking for. Engineering leads post about open roles on their teams.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;How to use X for job searching — step by step:&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;Step 1: Search for hiring signals in your niche. Try: “hiring [your role]”, “looking for a [role]”, “we’re building [domain] join us”&lt;/p&gt;
&lt;p&gt;Step 2: When you find a relevant post, reply to the tweet first don’t apply through a link. A reply to a founder’s hiring tweet gets seen. A cold form submission often doesn’t.&lt;/p&gt;
&lt;p&gt;Step 3: Follow and engage with the decision-maker before you need anything. Reply thoughtfully to their posts over a few weeks. By the time you DM, you’re not a stranger.&lt;/p&gt;
&lt;p&gt;Step 4: Send a direct, specific DM to the hiring manager. “Hi [name] I’ve been following your work on [X]. I saw you’re looking for a [role]. I’ve done [specific relevant thing]. Would it make sense to have a quick call?”&lt;/p&gt;
&lt;h2 id=&quot;2-wellfound-formerly-angellist-talent-startup-roles-that-are-actually-real&quot;&gt;2. Wellfound (formerly AngelList Talent) Startup Roles That Are Actually Real&lt;/h2&gt;
&lt;p&gt;Wellfound is built specifically for startup hiring. Listings here are almost entirely real startup founders paying for access have a direct incentive to keep listings current. Ghost jobs are far rarer because there’s accountability.&lt;/p&gt;
&lt;p&gt;The platform also shows things LinkedIn hides, approximate salary ranges, equity breakdown, company stage and funding status, actual team size. You can make an informed decision before applying which changes the whole dynamic.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Filter by funding stage and role to find companies actively building and hiring right now&lt;/li&gt;
&lt;li&gt;Read the founder’s note on each listing it tells you about culture and seriousness&lt;/li&gt;
&lt;li&gt;Apply directly, then find the founder or head of hiring on X or LinkedIn and send a brief, specific follow-up message&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;3-pallet--community-curated-job-boards&quot;&gt;3. Pallet- Community-Curated Job Boards&lt;/h2&gt;
&lt;p&gt;Pallet powers job boards run by specific professional communities, newsletters, Slack groups, creator audiences. Because these boards are curated by people with real reputations in their communities, quality is higher. A bad listing reflects on the community host, so there’s natural accountability.&lt;/p&gt;
&lt;p&gt;Find the Pallet board for your niche there are boards for product designers, growth marketers, developer advocates, climate tech roles, and dozens more. Small, targeted, and far less noisy than LinkedIn.&lt;/p&gt;
&lt;h2 id=&quot;4-professional-slack-communities--discordtg-servers&quot;&gt;4. Professional Slack Communities &amp;amp; Discord/TG Servers&lt;/h2&gt;
&lt;p&gt;Some of the best hiring conversations happening right now are in Slack and Discord communities and most people outside those communities have no idea. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://x.com/search?q=%23jobs&amp;amp;src=hashtag_click&quot;&gt;#jobs&lt;/a&gt; or &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://x.com/search?q=%23hiring&amp;amp;src=hashtag_click&quot;&gt;#hiring&lt;/a&gt; channels in active professional communities are where people post roles before they go anywhere public. They post there because they want a warm lead, not a cold applicant. Being a genuine, active participant contributing answers, sharing useful things, building a small reputation puts you in the running for roles that never get listed anywhere else..&lt;/p&gt;
&lt;h2 id=&quot;5-githubfor-engineers-the-profile-that-actually-matters&quot;&gt;5. GitHub — For Engineers, the Profile That Actually Matters&lt;/h2&gt;
&lt;p&gt;If you’re an engineer, your GitHub profile is more valuable than your LinkedIn profile. Hiring managers and technical leads look at your actual work: what you’ve built, how you’ve structured it, whether you write readable code, whether you contribute to open source.&lt;/p&gt;
&lt;p&gt;A strong GitHub profile with active contributions, good READMEs, and a project or two that demonstrates real skill will generate inbound recruiter interest in a way that keyword-optimising your LinkedIn profile simply won’t. Many open source projects also have community channels where maintainers post roles at their companies directly.&lt;/p&gt;
&lt;h2 id=&quot;6-personal-referralsstill-the-highest-converting-channel&quot;&gt;6. Personal Referrals — Still the Highest-Converting Channel&lt;/h2&gt;
&lt;p&gt;The data on this has been consistent for years, and it still holds.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://simplicant.com/blog/why-you-should-invest-in-employee-referrals-in-2026-5-key-benefits/&quot;&gt;A 2024 Aptitude Research report&lt;/a&gt; found that referrals convert to hires at &lt;strong&gt;4× the rate of non-referrals 30% conversion versus 7%.&lt;/strong&gt; &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.zippia.com/advice/employee-referral-statistics/&quot;&gt;Zippia’s 2026 analysis&lt;/a&gt; confirms referrals are 4× more likely to be offered a job than candidates applying through traditional job boards. And &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://staffinghub.com/hiring/company-pages-referrals-result-in-more-hires-recruiting-metrics-report/&quot;&gt;2026 data from StaffingHub&lt;/a&gt; found that referred candidates made up just 2% of applicants but accounted for 11% of hires making them 10× more likely to be hired.&lt;/p&gt;
&lt;p&gt;What this means practically: the highest-ROI activity in your job search is not applying to more jobs. It’s having more genuine conversations with people in your field.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Reach out to former colleagues for a referral also to genuinely reconnect and share what you’re working on&lt;/li&gt;
&lt;li&gt;Attend industry events, meetups, and conferences. In-person interactions still create stronger connections faster than digital ones&lt;/li&gt;
&lt;li&gt;Offer value before asking for anything. Share something useful, make an introduction for someone else, give a thoughtful piece of feedback&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;7-direct-company-applications--hiring-manager-follow-up&quot;&gt;7. Direct Company Applications + Hiring Manager Follow-Up&lt;/h2&gt;
&lt;p&gt;This sounds almost too obvious, but it genuinely works and most people skip the crucial second step. If there are 15 companies you’d genuinely love to work at, go to their actual career pages and apply directly. Your application doesn’t go through a job board algorithm. It goes to the hiring team.&lt;/p&gt;
&lt;p&gt;Then and this is the bit people miss, find the hiring manager for that role on LinkedIn or X. Not to send a formal application. To send a brief, human message:&lt;/p&gt;
&lt;p&gt;‘I just applied for your [role] through your site. I’ve been following what you’re building since [specific thing]. Happy to share more context if useful.’ That 30-second message has a higher conversion rate than 90% of the effort people put into LinkedIn.&lt;/p&gt;
&lt;h2 id=&quot;the-tactical-playbook-how-to-run-your-job-search-without-relying-on-linkedin&quot;&gt;The Tactical Playbook: How to Run Your Job Search Without Relying on LinkedIn&lt;/h2&gt;
&lt;p&gt;Let’s make this concrete. Here’s a structured approach that actually generates results.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Week One: Build Your Foundation&lt;/strong&gt;&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;Identify 20 companies you’d genuinely want to work at. Write them down. These are your targets.&lt;/li&gt;
&lt;li&gt;Set up alerts for each, Google Alerts for the company name, Crunchbase for funding news, their X/Twitter for hiring signals.&lt;/li&gt;
&lt;li&gt;Find 10 people on X or LinkedIn who are hiring managers, founders, or senior people at those companies. Follow them. Read what they write. Comment thoughtfully.&lt;/li&gt;
&lt;li&gt;Join two or three Slack or Discord communities relevant to your role and industry. Introduce yourself. Be helpful before you need anything.&lt;/li&gt;
&lt;li&gt;Optimise your GitHub, portfolio, or writing profile whichever is most relevant to what you do. This is the asset that compounds over time.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Week Two: Start Making Contact&lt;/strong&gt;&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;Look for X posts from your target companies with hiring signals. Reply before you DM.&lt;/li&gt;
&lt;li&gt;Apply directly to three to five companies through their career pages. Then send a brief follow-up to the relevant hiring manager.&lt;/li&gt;
&lt;li&gt;Reach out to three former colleagues not asking for referrals, just genuinely reconnecting and sharing what you’re looking for.&lt;/li&gt;
&lt;li&gt;Post something useful in one of your Slack communities. Answer a question. Share a resource. Be a person, not a job seeker.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Ongoing: What Consistency Actually Looks Like&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Two to three targeted direct applications per week, always followed by a human message to the hiring manager&lt;/li&gt;
&lt;li&gt;10 minutes daily on X checking for hiring signals in your niche&lt;/li&gt;
&lt;li&gt;One genuine community contribution per week a comment, a resource, a useful thread&lt;/li&gt;
&lt;li&gt;One coffee chat or call per week with someone in your field not a pitch, just a conversation&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The principle underneath all of this is the same: get as close to the actual decision-maker as possible, as early as possible, through the warmest channel available to you. Job boards LinkedIn or otherwise are cold and noisy. Direct relationships are warm and specific.&lt;/p&gt;
&lt;h2 id=&quot;the-uncomfortable-conclusion&quot;&gt;The Uncomfortable Conclusion&lt;/h2&gt;
&lt;p&gt;Here’s what I want to leave you with.&lt;/p&gt;
&lt;p&gt;LinkedIn is not going to fix itself. The incentive structure engagement metrics, ad revenue, premium subscription revenue rewards the current dysfunction more than it punishes it. Ghost jobs keep the listing numbers high. Fake profiles keep the membership numbers high. Hustle content keeps time-on-platform high. The platform is optimised for metrics that have very little to do with whether you actually get a job.&lt;/p&gt;
&lt;p&gt;So the move is not to rage-quit and cut yourself off from a platform where recruiters genuinely still search. The move is to deprioritise it. Use it as one narrow tool for one narrow job maintaining a findable profile and facilitating warm introductions. Then put the majority of your energy somewhere else.&lt;/p&gt;
&lt;p&gt;Build your presence where real conversations happen. Cultivate genuine relationships before you need them. Apply directly and follow up like a human. Get into the communities where good opportunities get shared before they go public.&lt;/p&gt;
&lt;p&gt;The job market is genuinely harder than it was three or four years ago. Layoffs have been widespread. Applications are at record highs. AI tools mean hiring managers are drowning in CVs, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.yahoo.com/news/employees-navigate-increasingly-soul-crushing-090000483.html&quot;&gt;Greenhouse reports&lt;/a&gt; that recruiter workload has increased 26% in the past year, with 91% of employees saying the current job market is challenging and 57% pointing to AI-driven competition. In that environment, doing what everyone else is doing spray-applying on LinkedIn and waiting is a strategy for frustration, not results.&lt;/p&gt;
&lt;p&gt;Being 10% more direct, 10% more specific, and 10% more human than the average applicant is enough to stand out. That’s the actual opportunity here.&lt;/p&gt;
&lt;p&gt;The jobs are out there. They’re just not always where the algorithm wants you to look.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;- If this resonated, pass it on to someone who’s been staring at an inbox that won’t move. They need to know it’s not them.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;Disclaimer: This article reflects personal opinion, observation, and cultural commentary. All statistics are sourced from publicly available research and reports. This is not intended to defame any individual, company, or platform.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>opinion</category><category>linkedin</category><category>jobs</category><category>job-hunting</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Your Dev Tool Isn’t Showing Up in AI Answers? Here’s Why and How to Fix It</title><link>https://soursen.org/articles/your-dev-tool-isnt-showing-up-in-ai-answers-heres-why-and-how-to-fix-it/</link><guid isPermaLink="true">https://soursen.org/articles/your-dev-tool-isnt-showing-up-in-ai-answers-heres-why-and-how-to-fix-it/</guid><description>Let’s be real, the way developers find tools has changed dramatically. They’re not Googling anymore, at least not the way they used to. They’re opening ChatGPT, Perplexity, or Claude and just asking. And if your product isn’t in the answer they get back? You weren’t even in the running.</description><pubDate>Sun, 31 May 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Let’s be real, the way developers find tools has changed dramatically. They’re not Googling anymore, at least not the way they used to. They’re opening ChatGPT, Perplexity, or Claude and just asking. And if your product isn’t in the answer they get back? You weren’t even in the running.&lt;/p&gt;
&lt;h2 id=&quot;the-shift-nobody-saw-coming-but-everyone-felt&quot;&gt;The Shift Nobody Saw Coming (But Everyone Felt)&lt;/h2&gt;
&lt;p&gt;Let me walk you through what this used to look like. A developer needs an observability platform. They open Chrome, search ‘best Prometheus alternatives for high-cardinality metrics,’ skim a few blog posts, check some GitHub star counts, and build a shortlist. That was the whole funnel. You optimized for it, it worked, everyone was happy.&lt;/p&gt;
&lt;p&gt;Now picture that same developer in 2026. They open Claude and type: ‘I’m running about 200 microservices on GCP, what observability stack would you go with and why?’ Thirty seconds later, they’ve got a synthesized answer comparing Grafana Cloud, Datadog, and Honeycomb complete with notes on pricing tiers, cardinality handling, and community reputation. They never opened your comparison article. They never saw your landing page. They got their answer and moved on.&lt;/p&gt;
&lt;p&gt;And this isn’t a niche behavior anymore. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://survey.stackoverflow.co/2025/ai/#sentiment-and-usage&quot;&gt;The Stack Overflow 2025 Developer Survey&lt;/a&gt; which polled over 49,000 developers across 177 countries found that 84% of respondents are using or planning to use AI tools in their development process, up from 76% the year before. Discovery has moved upstream, and it now happens inside a conversation with a model.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DBEXpxEy_1KTIt.webp&quot; alt=&quot;Illustration for the section “The Shift Nobody Saw Coming (But Everyone Felt)”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;680&quot; height=&quot;227&quot;&gt;&lt;/p&gt;
&lt;p&gt;Among professional developers, 51% use AI tools daily, indicating that AI has moved beyond experimental adoption into routine practice. While the survey doesn’t explicitly cite a “65% weekly or more” figure, it emphasizes that daily usage stands at 51% and overall adoption (including those planning to use AI) reaches 84%, underscoring rapid AI integration across the global developer community.. With the preponderance of developer tools integrating AI features, that number is only growing in 2026.&lt;/p&gt;
&lt;p&gt;So what do you do about it? That’s what AEO and GEO are about. Answer Engine Optimization and Generative Engine Optimization, two overlapping disciplines that define how you stay visible when the discovery layer is a language model, not a crawler.&lt;/p&gt;
&lt;p&gt;And here’s the thing &lt;strong&gt;if you’ve been producing quality technical content already, you’re not starting from scratch.&lt;/strong&gt; You just need to understand what changed, what stayed the same, and what moves you can make in the next 90 days to stop ceding ground to competitors who figured this out first.&lt;/p&gt;
&lt;h2 id=&quot;aeo-vs-geo-not-the-same-thing-but-close-enough-to-matter-together&quot;&gt;AEO vs. GEO: Not the Same Thing (But Close Enough to Matter Together)&lt;/h2&gt;
&lt;p&gt;You’ll see these two terms used interchangeably all the time, but they’re actually targeting different parts of the problem. For a deeper look at how each discipline is defined, the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Generative_engine_optimization&quot;&gt;Wikipedia entry on Generative Engine Optimization&lt;/a&gt; is a solid starting point. Here’s the practical difference:&lt;/p&gt;
&lt;p&gt;Answer Engine Optimization (AEO): &lt;strong&gt;AEO is about real-time retrieval.&lt;/strong&gt; When a developer asks Perplexity “what’s the best CI/CD tool for a Node.js monorepo,” Perplexity actively crawls the web, pulls content from recent sources, and synthesizes an answer. AEO is optimizing for that moment making sure your content is structured, fresh, and clear enough that an AI retriever can extract the relevant snippet and cite it in the answer.&lt;/p&gt;
&lt;p&gt;Generative Engine Optimization (GEO): &lt;strong&gt;GEO is about training data influence.&lt;/strong&gt; When a developer asks Claude “what’s a good alternative to Datadog for a small startup,” the model is pulling from what it learned during training and it’s not real-time retrieval. GEO means getting your brand, your product category associations, and your use-case anchors into the content that LLMs trained on.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CM8pjzyL_IfiVm.webp&quot; alt=&quot;Illustration for the section “AEO vs. GEO: Not the Same Thing (But Close Enough to Matter Together)”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;643&quot; height=&quot;301&quot;&gt;&lt;/p&gt;
&lt;p&gt;In practice, the best strategies run both tracks at once. Take a post like ‘Choosing a feature flag service for a multi-region SaaS product.’ You optimize it for AEO on your own site. FAQ schema, comparison table, clean structure, then distribute adapted versions across GEO surfaces: a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt; post, a Reddit thread contribution, an engineering newsletter write-up. One piece of content, two optimization tracks running in parallel.&lt;/p&gt;
&lt;h2 id=&quot;what-googles-february-2026-discover-core-update-actually-means&quot;&gt;What Google’s February 2026 Discover Core Update Actually Means&lt;/h2&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://developers.google.com/search/blog/2026/02/discover-core-update&quot;&gt;On February 5, 2026, Google released&lt;/a&gt; its first-ever Discover-specific core update. Google is now treating Discover as &lt;strong&gt;a separate ranking ecosystem&lt;/strong&gt; with its own quality criteria, totally distinct from Search.&lt;/p&gt;
&lt;p&gt;Three things changed that matter for developer content teams:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;1. Topic-by-Topic Expertise Evaluation:&lt;/strong&gt; Google’s systems now identify expertise per subject area. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.coremountainmedia.com/insights/google-discover-core-update-2026&quot;&gt;Core Mountain Media’s analysis&lt;/a&gt; puts it clearly, sites relying on sensational headlines and thin AI content saw drops of 30–60%, while sites with deep, consistent expertise gained ground. A regional tech news site with a deeply-covered DevSecOps vertical can outrank a large enterprise blog that published a single WebAssembly post.&lt;/p&gt;
&lt;p&gt;What this means for you? If you want to rank for ‘MLOps pipeline tooling,’ you need a real content cluster there. Not one overview post a cluster. Tutorials, integration guides, failure mode analyses, benchmark comparisons. Depth signals ownership. One article signals curiosity.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;&lt;strong&gt;Practical implication:&lt;/strong&gt;&lt;/em&gt; &lt;em&gt;Do a topic cluster audit right now. Pick one subject you want to own. Say, ‘edge computing for IoT.’ Can you point to a dozen pieces that build on each other? Architecture guides, provider comparisons, latency benchmarks, real-world case studies? If not, you’re not owning it yet, no matter how good your single best article is.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;strong&gt;2. Anti-Sensationalism Enforcement:&lt;/strong&gt; The updated Discover documentation is explicit about this &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ga.agency/en/google-discovery-core-february/&quot;&gt;GA Agency’s summary&lt;/a&gt; notes that Google now explicitly names ‘clickbait’ and ‘sensationalism’ as ranking demotions in updated guidance. Titles have to accurately reflect what’s actually in the piece.&lt;/p&gt;
&lt;p&gt;If your team has been A/B testing emotionally charged headlines, it’s time to stop. ‘How we cut p99 API latency from 800ms to 120ms using connection pooling’ will outperform ‘The shocking reason your API keeps failing under load’ for developers reading it and for AI systems interpreting it. Specific, descriptive, honest titles win on both fronts.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;3. Original, In-Depth Content Prioritization:&lt;/strong&gt; The update rewards content that brings something genuinely new. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.clickrank.ai/google-releases-discover-core-update-february/&quot;&gt;ClickRank’s guide to the February 2026 update&lt;/a&gt; confirms: original analysis, proprietary data, and first-hand experience signal the kind of depth Google’s systems are now prioritizing. Aggregation posts that collect existing coverage without adding a fresh perspective are algorithmically disadvantaged.&lt;/p&gt;
&lt;p&gt;The practical takeaway: your best content is written by engineers who ran the benchmarks, hit the edge cases, and have real opinions from building in production. Secondhand summaries are everywhere. Actual experience is scarce and that’s exactly what gets cited.&lt;/p&gt;
&lt;p&gt;This isn’t new advice but it’s now a ranking signal.&lt;/p&gt;
&lt;h2 id=&quot;what-actually-matters-in-geoaeo-and-what-youve-been-wasting-time-on&quot;&gt;What Actually Matters in GEO/AEO (And What You’ve Been Wasting Time On)&lt;/h2&gt;
&lt;p&gt;Not every tactic deserves equal attention. Let me give you a straight take on what’s moving the needle in 2026, what’s become less relevant, and what’s actively working against you.&lt;/p&gt;
&lt;h2 id=&quot;what-matters-most&quot;&gt;What Matters Most&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Topic-by-topic depth:&lt;/strong&gt; This is the single biggest shift in how content strategy works in 2026. Isolated articles about a tool category no longer build the kind of authority that drives AI citations. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://draft.dev/2026-developer-marketing-survey&quot;&gt;Draft.dev’s 2026 Developer Marketing Survey&lt;/a&gt; found that 58% of dev tool companies are increasing content investment this year, with budget shifting specifically toward depth-first strategies. The high-volume publishing play is losing ground.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Content freshness:&lt;/strong&gt; Recency matters to AI retrieval systems and in developer tooling, it matters double. An outdated tutorial that references a deprecated API or a renamed CLI flag doesn’t just fail to rank. It actively damages your credibility. A developer who follows your three-year-old Redis Streams walkthrough, hits a deprecation error on step three, and has to go figure it out themselves? That’s a trust hit that affects how they think about your whole brand.&lt;/p&gt;
&lt;p&gt;Google’s February 2026 update calls out timeliness explicitly as a Discover ranking signal. Build a refresh cadence into your content calendar at minimum once a year for technical pieces, quarterly for anything in fast-moving areas like AI tooling or cloud infrastructure&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Article structure:&lt;/strong&gt; Here’s something that’s easy to overlook, AI retrieval systems favor pages that are built for extraction. A &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://machinerelations.ai/research/content-structure-ai-citation-rates-2026&quot;&gt;March 2026 study from researchers at the University of Tokyo and University of Tsukuba&lt;/a&gt; found that structural optimization alone independent of content quality produces a consistent 17.3% improvement in AI citation rates. Practically, this means:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Skimmable (summary or bullet points at the top)&lt;/li&gt;
&lt;li&gt;Structured (tables and matrices for comparisons)&lt;/li&gt;
&lt;li&gt;Self-explanatory (each section makes sense independently)&lt;/li&gt;
&lt;li&gt;Conclusive (key takeaways or FAQs at the bottom)&lt;/li&gt;
&lt;li&gt;Well linked (internally and externally)&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Good news: none of this is AI-specific. It makes your content better for human readers and for traditional SEO too. It’s just that the stakes are higher now for AI systems, the difference between a well-structured article and a flowing narrative essay can determine whether you get cited at all.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;External brand mentions:&lt;/strong&gt; Here’s a mental model shift worth making, in a world where LLMs weight distributed brand signals over concentrated link authority, your content strategy can’t stop at your own domain. Research cited in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mersel.ai/generative-engine-optimization&quot;&gt;Mersel AI’s GEO guide&lt;/a&gt; references the original Princeton GEO study finding that citations boosted AI visibility by +30% and quotations by +41%. A mention of your tool in a well-read engineering newsletter even without a link contributes to LLM training data in ways that a tenth backlink from a low-traffic blog simply doesn’t.&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Organic references in online engineering newsletters with archived web versions&lt;/li&gt;
&lt;li&gt;Posts from your own team and early customers on developer-first platforms like Hashnode and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;Genuine participation in relevant subreddit discussions and not promotional drops or links, but actual technical contributions that happen to mention your product&lt;/li&gt;
&lt;li&gt;User-generated case studies and retrospectives hosted on personal or company blogs&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;And here’s the leveling opportunity: newer companies can catch up faster through distributed brand presence than they ever could through backlink campaigns. If you’ve been building in public, shipping technical content to community platforms, and cultivating users who write about their experience you’re already accumulating LLM mind-share in ways that a company with a bigger domain authority might not be.&lt;/p&gt;
&lt;h2 id=&quot;what-doesnt-matter-as-much-anymore&quot;&gt;What Doesn’t Matter as Much Anymore&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Domain Age or Authority Rating:&lt;/strong&gt; Domain Rating, Moz DA, and similar metrics have always been proxies useful approximations. In 2026, they’re even less predictive. The February 2026 Discover update evaluates expertise per subject, which means a newer domain with deep, genuine coverage of a specific developer tool category can outrank a well-established domain that’s only touched that subject superficially. Age is no longer a moat. Depth is.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Traditional Backlinking Campaigns:&lt;/strong&gt; Link-building as a standalone play has diminishing returns in the AI discovery context. LLMs aren’t running PageRank on your inbound link graph they’re looking at brand presence, topical association, and community sentiment across a distributed set of crawled surfaces. A genuine recommendation in a relevant forum thread carries more weight in model training data than a link from an off-topic guest post.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;High-Volume, Pure AI-Generated Content:&lt;/strong&gt; This one is worth being blunt about. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://draft.dev/2026-developer-marketing-survey&quot;&gt;Draft.dev’s 2026 survey&lt;/a&gt; found that 96% of developer marketing teams have used AI generation in their workflow but only 7% describe the results as ‘very useful.’ Practitioner-written content generates 4.8× more engagement than unreviewed AI output. The problem isn’t using AI tools most high-performing teams do. It’s removing domain expertise from the authoring process entirely.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Gated content without an ungated version:&lt;/strong&gt; If LLMs can’t index it, it doesn’t exist for AI-powered discovery.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;&lt;strong&gt;PATTERNS THAT HURT YOU — GOOGLE AND LLMS BOTH PENALIZE THESE:&lt;/strong&gt;&lt;/em&gt; &lt;em&gt;Filler phrases (“with that in mind…”, “at the end of the day”) · Headlines that promise more than you deliver · Intros that just restate the title · Hub pages that only link to other pages · Every section starting with the same template sentence · Gated content with no crawlable preview&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;the-numbers-behind-the-shift-what-the-research-actually-shows&quot;&gt;The Numbers Behind the Shift: What the Research Actually Shows&lt;/h2&gt;
&lt;p&gt;Let’s ground this in actual data. Not the kind that gets dressed up for strategy decks, but the kind that should actually change where you put your time and budget.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CWnX4J6j_Z2eBbVN.webp&quot; alt=&quot;Illustration for the section “The Numbers Behind the Shift: What the Research Actually Shows”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;580&quot; height=&quot;215&quot;&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-ai-content-paradox&quot;&gt;The AI Content Paradox&lt;/h2&gt;
&lt;p&gt;Here’s the stat that should genuinely give you pause: according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://draft.dev/2026-developer-marketing-survey&quot;&gt;Draft.dev’s 2026 Developer Marketing Survey&lt;/a&gt;, 96% of developer marketing teams have used AI generation in their content workflow. But only 7% say the results are ‘very useful’ and practitioner-authored content generates 4.8× more engagement than unreviewed AI output.&lt;/p&gt;
&lt;p&gt;So why does everyone keep doing it? Because it feels productive. Volume is up, turnaround is faster, the calendar looks full. But the performance gap is real and it’s growing. The companies winning in 2026 are using AI to accelerate a workflow that still starts and ends with genuine domain knowledge.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.DxQi5-ZG_2jCczK.webp&quot; alt=&quot;Illustration for the section “The AI Content Paradox”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;642&quot; height=&quot;207&quot;&gt;&lt;/p&gt;
&lt;h2 id=&quot;where-your-brand-needs-to-live-and-how-to-get-it-there&quot;&gt;Where Your Brand Needs to Live (And How to Get It There)&lt;/h2&gt;
&lt;p&gt;One of the most important mindset shifts in AEO/GEO is this, the surfaces that matter are no longer limited to your own domain. LLMs train on a distributed ecosystem of crawled content. Real-time retrieval engines index all of it. Your content strategy has to reflect the full picture of where developer conversations actually happen.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.Bd4fkNyh_1k9ii2.webp&quot; alt=&quot;Illustration for the section “Where Your Brand Needs to Live (And How to Get It There)”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;640&quot; height=&quot;351&quot;&gt;&lt;/p&gt;
&lt;p&gt;The opportunity here is real, especially for newer companies. An established player might have a strong backlink profile but thin community presence. A startup that ships genuine technical content on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt;, participates honestly in relevant subreddits, and builds an audience of users who write publicly about their experience? That company can close the LLM mind-share gap faster than traditional SEO ever allowed.&lt;/p&gt;
&lt;h2 id=&quot;the-5-content-types-that-win-in-the-ai-era&quot;&gt;The 5 Content Types That Win in the AI Era&lt;/h2&gt;
&lt;p&gt;Not all content formats get treated equally by AI retrieval systems. These five consistently outperform and there are concrete reasons why each one works.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;1. Comparison and Roundup Content:&lt;/strong&gt; Comparison content is one of the most-cited formats in AI-generated recommendations, and it’s not hard to understand why. A post titled ‘Kafka vs. Pulsar vs. Redpanda: picking a streaming platform for an event-driven microservices architecture’ gives AI systems exactly what they need a structured evaluation across defined criteria, with clear guidance about which option fits which constraints.&lt;/p&gt;
&lt;p&gt;The difference between comparison content that gets cited and content that doesn’t is substantiation. ‘We found X performed better’ is a claim. ‘We processed 2M events per second on a three-node cluster and saw X behavior under backpressure’ is evidence. AI systems, like developers, trust the latter.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;2. Tutorials with Real-World Scenarios:&lt;/strong&gt; Well-crafted tutorials using production-realistic scenarios are genuinely high-value training data for AI coding assistants. A tutorial on implementing distributed rate limiting with Redis Lua scripts in a multi-region setup with working, tested code and honest edge case discussion is far more useful to a model than a ‘getting started’ example using localhost and toy data.&lt;/p&gt;
&lt;p&gt;Freshness matters more here than almost anywhere else. Outdated API references in tutorials are one of the fastest ways to lose developer trust and drop out of AI citation. Make tutorial refresh a standing item in your content calendar.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;3. Industry-Specific Guides:&lt;/strong&gt; AI assistants adapt their answers to context. ‘What observability tooling would you recommend for a Series A healthtech company handling patient data?’ gets a different answer than the same question without that context and rightly so. Content that connects your tool’s capabilities to the specific constraints of a given vertical: the regulatory requirements in fintech, the latency sensitivity in gaming infrastructure, the compliance overhead in healthcare positions you to capture those contextually-rich queries.&lt;/p&gt;
&lt;p&gt;The practical takeaway: write one piece for the MLOps engineer at a genomics company, not one piece for every ML engineer everywhere. Specificity is a feature, not a limitation.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;4. Original Analysis and Proprietary Data:&lt;/strong&gt; The 2026 emphasis on originality isn’t abstract it shows up clearly in citation rates. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mersel.ai/generative-engine-optimization&quot;&gt;The original GEO research&lt;/a&gt; found that content containing statistics boosted AI visibility by +32%, and cited sources by +30%. Content that has data no one else has is structurally irreplaceable latency benchmarks, customer survey results, honest failure postmortems. Whatever form it takes: publish it ungated, make it crawlable, and don’t bury the numbers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;5. Decision Frameworks:&lt;/strong&gt; Content that helps a developer make a specific, bounded decision outperforms content that describes a topic at length. The structure that works: establish the decision context, spell out the criteria that should drive the choice, then give clear ‘choose X if you need…’ guidance with concrete conditions attached.&lt;/p&gt;
&lt;p&gt;This format is practically designed for AI extraction. Models love to surface decision frameworks because they answer the real question behind most tool queries which isn’t ‘tell me about CI/CD platforms,’ it’s ‘which one should I actually use.’ Content that answers the real question gets cited. Content that dances around it doesn’t.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-06.Bx8NBtt2_pwT4V.webp&quot; alt=&quot;Illustration for the section “The 5 Content Types That Win in the AI Era”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;606&quot; height=&quot;335&quot;&gt;&lt;/p&gt;
&lt;h2 id=&quot;how-to-adapt-without-starting-from-scratch-your-90-day-playbook&quot;&gt;How to Adapt Without Starting From Scratch: Your 90-Day Playbook&lt;/h2&gt;
&lt;p&gt;If your team has been investing in technical content, you’re already holding the core asset. AEO and GEO don’t ask you to throw out what you’ve built, they ask you to extend it for a discovery landscape that now includes AI systems. Here’s a concrete 90-day sequence to do that without rebuilding from zero.&lt;/p&gt;
&lt;h2 id=&quot;days-130-audit-and-foundations&quot;&gt;Days 1–30: Audit and Foundations&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Split your Discover and Search analytics in Google Search Console. They’re separate channels now and need to be tracked that way.&lt;/li&gt;
&lt;li&gt;Map your content against topic clusters. Figure out which subjects you genuinely own with depth, and which are covered by a single article floating on its own with no supporting structure.&lt;/li&gt;
&lt;li&gt;Audit your top 20 posts. Look specifically for missing summaries, absent comparison tables, and no FAQ sections. Prioritize pieces that already have traffic structure improvements compound fastest there.&lt;/li&gt;
&lt;li&gt;Pull your ten highest-traffic tutorials. Flag anything with outdated syntax, deprecated API calls, or library versions that have moved on. These are your first refresh queue.&lt;/li&gt;
&lt;li&gt;Add FAQ schema to your top comparison and tutorial pages. It’s a one-time effort that keeps paying off with every retrieval event.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;days-3160-content-and-structure&quot;&gt;Days 31–60: Content and Structure&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Write or commission at least two comparison pieces framed as decision guides, not feature lists. ‘Choosing between Temporal and Apache Airflow for a financial data pipeline’ beats ‘Temporal vs. Airflow: which is better’ for readers and for AI systems. Publish on your site first, then adapt for &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Add TL;DR summaries to your top ten existing articles. Lead with the conclusion. AI systems index summary sections preferentially, and readers who skim appreciate it just as much.&lt;/li&gt;
&lt;li&gt;Dig out any gated whitepapers or reports with proprietary data. Publish ungated versions of the data-heavy sections. Gated content contributes nothing to LLM training data or real-time retrieval.&lt;/li&gt;
&lt;li&gt;Pick three to five industry verticals where your product has a real, specific advantage. Brief one content piece per vertical that speaks directly to the constraints and context of that industry.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;days-6190-distribution-and-brand-mentions&quot;&gt;Days 61–90: Distribution and Brand Mentions&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Push adapted versions of your comparison content to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt;, Hashnode, and two or three relevant subreddits. ‘Adapted’ is the key word a direct cross-post adds no value to those communities. Reframe the angle for each audience.&lt;/li&gt;
&lt;li&gt;Identify ten to fifteen engineers and technical writers in your ecosystem who publish regularly on crawled platforms. Build actual relationships before you need the mentions. Transactional outreach is obvious and it doesn’t work.&lt;/li&gt;
&lt;li&gt;Review your GitHub README and public docs. Both should clearly explain your use case, name the specific problems you solve, and reference alternatives honestly. Developers and AI systems both value directness over marketing language.&lt;/li&gt;
&lt;li&gt;Set up brand mention tracking Mention, Brand24, or manual Reddit and Hacker News searches. Establish a baseline this month. Watch the trend from here.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;the-companies-who-win-this-are-already-starting&quot;&gt;The Companies Who Win This Are Already Starting&lt;/h2&gt;
&lt;p&gt;Here’s the honest version of where things stand, the first-mover window in AEO and GEO hasn’t closed, but it’s closing. The dev tool companies that will hold the strongest positions in AI-generated recommendations two or three years from now are making content investments today that compound building genuine topic clusters, distributing real expertise across the platforms developers trust, and earning organic mentions through actual product quality and community participation.&lt;/p&gt;
&lt;p&gt;The companies that will lose ground are the ones still running a 2022 playbook, treating AI generation as a content factory, publishing at volume without substance, and acting as if their own blog is the only discovery surface that matters. That approach still produces some results. It won’t in two years.&lt;/p&gt;
&lt;p&gt;The good news is that the fundamentals haven’t moved. Expert-written content that solves real problems, backed by evidence from actual usage, is the foundation for SEO, for AEO, and for GEO. Everything else in this document is additive: better structure, smarter distribution, fresher content, broader brand presence across the platforms where developers actually spend time.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Start with the audit. The rest follows.&lt;/strong&gt;&lt;/p&gt;
&lt;h2 id=&quot;quick-start-summary&quot;&gt;Quick-Start Summary&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;Keep writing substantive, practitioner-level content. The core stuff hasn’t changed.&lt;/li&gt;
&lt;li&gt;Refresh your tutorials and API docs regularly. Outdated code is worse than no content.&lt;/li&gt;
&lt;li&gt;Add structure everywhere: FAQs, comparison tables, section summaries.&lt;/li&gt;
&lt;li&gt;Go deep on fewer topics rather than shallow on many.&lt;/li&gt;
&lt;li&gt;Show up on more platforms: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://dev.to/&quot;&gt;Dev.to&lt;/a&gt;, Reddit, Hashnode, newsletters, partner blogs.&lt;/li&gt;
&lt;li&gt;Cut the fluff. If a sentence doesn’t add anything, delete it.&lt;/li&gt;
&lt;li&gt;Track Discover and Search as separate channels in Google Search Console.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;em&gt;Found this useful? Share it with your developer marketing or DevRel team. The playbook works but it only works if you start.&lt;/em&gt;&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Developer tools</category><category>aeo</category><category>geo</category><category>seo</category><category>ai</category><category>generative-ai-tools</category><author>hello@soursen.org (Shivashish Yadav)</author></item></channel></rss>