<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Tim Chosen on Medium]]></title>
        <description><![CDATA[Stories by Tim Chosen on Medium]]></description>
        <link>https://medium.com/@timchosen?source=rss-a3aaf36b7919------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*JY-2leJbkgdC4gV7ctNxNA.jpeg</url>
            <title>Stories by Tim Chosen on Medium</title>
            <link>https://medium.com/@timchosen?source=rss-a3aaf36b7919------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 21:11:10 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@timchosen/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A Fake Technical Interview Tried to Get Me to Run a Malicious Repository]]></title>
            <link>https://timchosen.medium.com/a-fake-technical-interview-tried-to-get-me-to-run-a-malicious-repository-631af261b1aa?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/631af261b1aa</guid>
            <category><![CDATA[web3]]></category>
            <category><![CDATA[blockchain-technology]]></category>
            <category><![CDATA[job-hunting]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[linkedin]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 12:52:50 GMT</pubDate>
            <atom:updated>2026-11-18T12:52:50.143Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The warning signs where there, but it wasn’t just the GitHub link. It was the insistence that I open and run the project before I understood what it did.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VHgoDjyUo-kSuVNPLBtVag.png" /><figcaption>The “recruiter” from LinkedIn that reached out</figcaption></figure><p>I hadn’t interviewed for a job in almost three years. So when I started looking again a few weeks ago, AI and changing recruitment practices made me feel like I was fresh out of Uni, learning how to get hired all over again.</p><p>That felt strange because In 2017, I built an early AI-powered CV platform designed to help applicants get past applicant tracking systems. Users chatted with a CV bot to create a CV and we tailored their CVs to the language those ATS systems looked for. I’d also spent the past three years on the hiring side, so I <em>thought</em> knew what hiring managers were looking for.</p><p>Even so, my search brought the familiar scams like the “confidential job openings” that seemed like a perfect fit, provided I paid a few hundred pounds for a CV upgrade. Those were easy to ignore. The attempt I want to talk about was different. It tried to get me to run malicious code on my computer.</p><p>I joined a call that appeared to be about a technical leadership opportunity at an AI company. The person introduced himself as a project director, asked about my engineering background, and described an AI-agent marketplace with a working MVP. He then sent me a GitHub repository and suggested we review the project together.</p><p>That would not, on its own, have been unusual. Engineers routinely review code during interviews. What concerned me was the direction of the conversation when he repeatedly asked me to share my <strong>entire screen</strong>, and insisted I open the project in Cursor even when I said I wanted to use code editor, clone the repository, and work through its setup. When I asked what he actually wanted me to do with the repo, I didn’t get a clear technical task.</p><p>So I stopped to inspect the code before running it and I made sure he heard me say, let me get claude to check the code for any malicious code.</p><p>The moment I said, “I’m scanning the code if there’s any malicious stuff in there. Before I run it,” the person on the call dropped off.</p><h3>What I found in the repository</h3><p>Once he dropped off, I became even more interested and went ahead to review the source code and found <strong>two apparent execution paths</strong> concealed inside an application presented as an AI-agent marketplace.</p><p>The first was in .vscode/tasks.json. The tasks were configured with runOn: &quot;folderOpen&quot; and a hidden terminal. On macOS, the command was designed to download a script from an external endpoint and pipe it into bash; the repository contained corresponding patterns for Linux and Windows. The command text was padded with whitespace, making the consequential part easier to overlook. A workspace setting also hid the .vscode folder from the editor’s file tree.</p><p>The second path was npm install. A prepare script in package.json started a Node server. In the code I reviewed, loading that server would POST process.env—the process’s environment variables—to an external endpoint. It would then execute JavaScript returned by that endpoint with access to Node’s require.</p><p>Those are not ordinary requirements for viewing an AI-agent marketplace. They are mechanisms for sending environment data off-machine and running code supplied by a remote server. I <strong>did not download the second-stage payloads</strong>, so I cannot say what they contained.</p><h3>This pattern has succeeded before</h3><p>The broader tactic of using a purported job opportunity to persuade a developer to run attacker-supplied code is has seen a steady increase. In November 2026, Binance disclosed that a fake recruitment process was part of the intrusion that led to a reported <strong>$6.3 billion crypto theft</strong>. Reuters reported that, after interviews for what appeared to be a legitimate role, a Binance software engineer downloaded a malicious code repository. The attackers subsequently used the compromised machine and other access to reach wallet infrastructure. That is a comparison of <strong>method</strong>, not a claim that the person who contacted me was connected to that attack.</p><p>There is an even closer technical parallel. Reporting on Microsoft’s research into fake developer-assessment repositories describes malicious VS Code workspace tasks using runOn: &quot;folderOpen&quot; to retrieve attacker-controlled code when a developer opens and trusts a project. The same research documented other execution paths embedded in apparently legitimate development workflows.</p><p>ReversingLabs has also documented fake recruiter campaigns targeting crypto developers with interview tasks that led to malicious dependencies. Its research illustrates why a repository can look like a normal coding exercise while the dangerous behaviour sits in setup steps or packages rather than the feature code a candidate expects to review.</p><p>These cases do not prove who created the repository I received. They do show that the request to “just open the project” or “just install the dependencies” can be the attack, not merely preparation for an interview.</p><h3>The question that changed the call</h3><p>I asked what the exercise was for. I wanted to know what I was meant to assess or build before running somebody else’s code on my machine. I did not receive a clear answer, and the call ended when I said I was checking for malicious code.</p><p>My takeaway for other developers is simple: <strong>treat an interview repository with the same caution as any other untrusted code</strong>. Inspect workspace automation and package scripts before opening the folder as a trusted editor workspace or running an install. If execution is genuinely necessary, establish the purpose first and use an isolated environment.</p><p>A legitimate interviewer should be able to explain the task and give you time to review what you are being asked to run.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=631af261b1aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Weekly Ledger]]></title>
            <link>https://medium.com/the-ai-weekly-ledger/the-ai-weekly-ledger-d842043a7c5b?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/d842043a7c5b</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[european-union]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 12:17:18 GMT</pubDate>
            <atom:updated>2026-11-18T12:17:18.371Z</atom:updated>
            <content:encoded><![CDATA[<p>Tuesday 17 November 2026</p><p>This week Anthropic began pitching investors on a public share sale, and Google released a cheaper AI model for coding and automated tasks along with two research results, one in DNA analysis and one in software security. In Europe, the Commission’s science advisers asked for stronger AI rules and independent testing, and the Commission published a draft form for checking how AI affects people’s rights.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eTT-UYaqxzTb4lqj" /><figcaption>Photo by <a href="https://unsplash.com/@brechtcorbeel?utm_source=medium&amp;utm_medium=referral">Brecht Corbeel</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Anthropic starts selling shares to the public in a sale that could raise about $60 billion</h3><p>On 10 November, Anthropic, the maker of the Claude AI assistant, <a href="https://www.anthropic.com/news/roadshow-launch">opened its pitch to investors</a> for a sale of shares on the Nasdaq stock exchange under the symbol ANT. The sale totals 825 million shares. Anthropic is selling 801,351,354 of them and existing shareholders are selling 23,648,646, with no money from those going to the company. The expected price is $70 to $78 a share, which would bring Anthropic roughly $56 billion to $62.5 billion before fees. Goldman Sachs and Morgan Stanley are the banks handling the sale, and they can buy up to 123,750,000 more shares within 30 days.</p><p>An initial public offering, or IPO, is the first time a company sells shares to anyone who wants them, rather than only to private investors. Anthropic filed its disclosure document with the US Securities and Exchange Commission, the main American market regulator, on 26 October. The roadshow is the round of meetings in which executives persuade large investors to buy before the final price is set.</p><p>Investors should read the risk section of that filing before deciding, since it is the company’s own account of what could go wrong. Founders and operators who build on Claude should expect the company’s finances to face public scrutiny once trading starts.</p><h3>Research</h3><h3>Google’s AI checks the core encryption code of the Signal messaging app</h3><p>Google DeepMind said on 10 November that its <a href="https://deepmind.google/blog/introducing-alphaproof-code/">AlphaProof Code system</a> verified the core encryption code of the Signal protocol, the security system behind the Signal messaging app, down to machine instructions. People decided what to prove and several AI models wrote the mathematical proofs, which took four weeks. Human reviewers checked about 1 percent of the code, the parts the AI flagged. As a test, the team put back two bugs that had been fixed earlier, and the system produced machine-checked proof that each version was insecure. Google DeepMind says this is the first time AI has verified a widely used encryption protocol at this depth, and it has published the proofs openly.</p><p>Security teams that depend on encryption should watch whether this method becomes available for their own code. For now it is a research result, not a product.</p><h3>Google releases a DNA model that finds nearly twice as many gene-changing variants</h3><p>Google DeepMind released <a href="https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/">AlphaGenome 2</a> on 11 November, a model that predicts how changes in DNA affect specific types of human cells. Google says it finds nearly twice as many variants that change how active a gene is as the earlier version. On a test set covering 28 immune cell types, it scored 30 percent against 16 percent for the earlier model at finding variants that cause a change, and 43 percent against 23 percent at predicting the direction of the effect. It is available now for non-commercial research through Google’s software interface, with a private preview for commercial research on Google Cloud.</p><p>Google states the model is not approved for clinical or diagnostic use. Biotech teams can test it on research questions today but should not build patient-facing tools on it.</p><h3>Products and funding</h3><h3>Google ships a faster, cheaper AI model for coding and automated tasks</h3><p>Google launched <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/">Gemini 4 Hydrogen</a> on 12 November, a model built for software agents and everyday coding. It costs $1.50 per million input tokens and $7.50 per million output tokens, with a context window of 1 million tokens (tokens are the small chunks of text that AI models read and write). In Google’s own tests it completed 57 percent of tasks on a command-line coding benchmark, against 18.2 percent for its earlier Gemini 3.8 Flash. It is rolling out in the Gemini developer tools, the Gemini app and Google Search.</p><p>Operators running AI agents should price it against their current model, and test it on their own work, since the results above come from Google.</p><h3>British chip start-up Fractile raises a round worth over £500 million with government backing</h3><p>Fractile, a Bristol and London company designing AI chips, <a href="https://www.gov.uk/government/news/sovereign-ai-and-british-business-bank-invest-in-british-firm-developing-the-next-generation-of-ai-hardware">secured a Series C round</a> worth over £500 million that values it at over £5 billion, the UK government said on 11 November. The investors include the government’s Sovereign AI Venture Fund and the British Business Bank, along with CapitalG, Lightspeed and Redpoint. The government did not disclose how much it put in. Fractile says its chips are meant to make AI hardware faster, more energy efficient and cheaper to run.</p><p>Buyers of AI computing should treat Fractile as a supplier to watch rather than one to plan around until its chips ship. Investors can read the round as a sign of how much public money the UK is willing to put behind home-grown AI hardware.</p><h3>Regulation and policy</h3><h3>The European Commission’s science advisers tell it to keep AI rules strong and test AI models itself</h3><p>On 16 November two advisory bodies to the European Commission, the Group of Chief Scientific Advisors and the Group on Ethics in Science and New Technologies, <a href="https://research-and-innovation.ec.europa.eu/news/all-research-and-innovation-news/ai-requires-rigorous-safeguards-and-rules-keep-pace-technological-advances-experts-tell-european-2026-11-16_en">issued a joint statement</a> on governing AI. They say existing guardrails no longer suffice, that outside efforts to delay or weaken the law could compromise safeguards, and that the EU should build its own capacity to evaluate AI models rather than leave that to the companies that make them. They also want the pace of AI development and release matched to how dangerous a system could be. The Commission’s own <a href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai">AI Act page</a> says it plans EU evaluation capacity for AI models by 2027.</p><p>AI companies selling in Europe should prepare for testing by outside evaluators, not only their own safety reports. The statement is advice, not law, but it adds weight to those arguing against delay.</p><h3>Europe publishes a draft form for companies to check how their AI affects people’s rights</h3><p>On 16 November the European Commission <a href="https://digital-strategy.ec.europa.eu/en/news/commission-seeks-feedback-draft-template-fundamental-rights-impact-assessment-under-ai-act">asked for feedback</a> on a draft questionnaire that certain organisations will have to complete before using high-risk AI systems under the AI Act, the EU’s main AI law. The Commission’s AI Office wrote it with the EU Agency for Fundamental Rights. Feedback is open until 15 January 2027, and organisations can volunteer to test the draft on one of their own systems by 4 December 2026. The duty to complete the assessment applies from 2 December 2027.</p><p>Companies that use AI in areas such as hiring or education should read the draft now and consider volunteering, since testers get a say in the final form.</p><h3>Next week</h3><p>Monday 30 November is the date by which researchers should submit their use case to join Google DeepMind’s AlphaGenome 2 community call.</p><p>Friday 4 December is the deadline to volunteer to test the European Commission’s draft rights-assessment form with the AI Office.</p><p>Friday 15 January 2027 is when public feedback on that draft form closes.</p><p>That is the issue. See you next Tuesday. — The AI Weekly Ledger</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d842043a7c5b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-ai-weekly-ledger/the-ai-weekly-ledger-d842043a7c5b">The AI Weekly Ledger</a> was originally published in <a href="https://medium.com/the-ai-weekly-ledger">The AI Weekly Ledger</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Weekly Ledger]]></title>
            <link>https://medium.com/the-ai-weekly-ledger/the-ai-weekly-ledger-40d0ebef5686?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/40d0ebef5686</guid>
            <category><![CDATA[ai-security]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 12:13:22 GMT</pubDate>
            <atom:updated>2026-11-18T12:13:22.553Z</atom:updated>
            <content:encoded><![CDATA[<p>Tuesday 10 November 2026</p><p>Attackers are now using AI programs to break into company cloud accounts faster than a security team can react, and Microsoft has published the evidence. This week also brought a new American task force to prosecute AI crime, a European court opinion that could change how chatbots use news articles, and tighter British rules on police facial recognition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U8uJynkCacJNFiQD" /><figcaption>Photo by <a href="https://unsplash.com/@flyd2069?utm_source=medium&amp;utm_medium=referral">FlyD</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Attackers used an AI program to turn a leaked password into stolen company data in five minutes</h3><p>Microsoft’s security researchers said on 5 November that criminals are using an automated hacking program called ARTEX to break into company cloud accounts. <a href="https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/">Their report</a> describes a program that works through the steps of an attack on its own, without a person directing each move.</p><p>The method is simple. Developers sometimes leave passwords for their applications inside code they publish on GitHub, the website where most of the world’s software is stored and shared. ARTEX searches that public code for these passwords, then uses them to sign in to the victim’s accounts on Microsoft’s cloud services, including Azure and Microsoft 365.</p><p>Microsoft says the activity began in August 2026 and that the main wave of sign-ins started on 28 October. In one case the program went from finding a password to copying data out of a company’s vault of stored credentials in five minutes.</p><p>A human attacker might take hours or days to do the same work, which gave defenders time to spot a leak and change the password. That margin has gone. Any company with code on GitHub should scan it for leaked passwords this week and replace any it finds. Microsoft also recommends moving from stored passwords to managed identities, where the cloud provider issues short-lived credentials automatically and no password exists to leak.</p><h3>Research</h3><h3>Britain’s cyber security agency expects an AI-powered attack on essential services by 2028</h3><p>The National Cyber Security Centre, the government body that defends the United Kingdom against online attacks, published its annual review on 3 November. <a href="https://www.ncsc.gov.uk/news/cyber-chief-uk-facing-ai-powered-attacks-on-critical-infrastructure-in-two-years">The agency says</a> it is highly likely that malicious software driven by AI will be used by 2028 in a ransomware attack, where criminals lock an organisation’s systems and demand payment, against services such as energy, water or health. Its chief executive, Dr Richard Horne, urged organisations to spend on resilience and warned against buying technology on the cheap. The government announced £60 million for a national Cyber Shield and £70 million for a new response centre.</p><p>Operators of essential services and their suppliers should expect customers and regulators to ask for proof of security spending. Security companies should watch for the contracts that follow the new funding.</p><h3>Products and funding</h3><h3>Mira Murati’s company released a new AI model and plans to make it free to download</h3><p>Thinking Machines Lab, founded by former OpenAI executive Mira Murati, <a href="https://thinkingmachines.ai/news/inkling-2-small/">released Inkling 2 Small</a> on 5 November. The model can read images as well as text and is available now through Tinker, the company’s own platform. It has 594 billion parameters, the adjustable values that store what a model has learned, and uses 17 billion of them for each piece of text it produces, which keeps running costs down. The company says it will publish the full model under the Apache 2.0 licence, which permits free commercial use, once outside safety testing is complete.</p><p>Companies that want to run a capable model on their own servers, without sending data to a supplier, should plan to test it when the download appears. The release depends on the safety tests, so treat the timing as open.</p><h3>Regulation and policy</h3><h3>The United States set up a team of prosecutors to pursue crimes committed with advanced AI</h3><p>The Department of Justice <a href="https://www.justice.gov/opa/pr/department-justice-launches-super-intelligence-task-force">announced a Super Intelligence Task Force</a> on 5 November. It is led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss. Its stated aim is to stop foreign governments, international criminal groups and fraudsters from misusing advanced AI. Attorney General Todd Blanche said the department will prosecute unlawful use.</p><p>Companies that sell AI tools, especially ones that could be used for fraud or impersonation, should expect more federal investigations and keep records of how they screen customers. The announcement creates no new rules, so existing criminal law applies.</p><h3>A senior adviser to Europe’s top court said AI chatbots can break the law when they reuse news articles</h3><p>On 3 November, Advocate General Szpunar gave <a href="https://curia.europa.eu/site/upload/docs/application/pdf/2026-11/cp260155en.pdf">his opinion</a> in the first generative AI case before the Court of Justice of the European Union, the bloc’s highest court. A Hungarian publisher is challenging Google over how its Gemini chatbot retrieves and reuses press articles. He said that training an AI model on articles is generally allowed under an existing exception for automated text analysis. A chatbot’s answer can still infringe a publisher’s rights if the article was reached unlawfully, for example by getting past a paywall, or if the answer copies the article closely enough that the reader no longer needs to visit the original. The opinion does not bind the judges, who will rule later.</p><p>AI companies operating in Europe should review how their products fetch and summarise news before the judgment arrives. Publishers now have a stronger hand in licensing talks.</p><h3>Britain will require police to follow a specific law when they use facial recognition</h3><p>The Home Office, the ministry responsible for policing, <a href="https://www.gov.uk/government/news/stronger-laws-for-police-using-facial-recognition-technology">said on 4 November</a> that it will bring in a law requiring police use of facial recognition to be proportionate and based on intelligence. An independent watchdog will oversee forces and, in extreme cases, can ask a court to stop a force using the technology. Independent testing for bias by age, gender and race becomes mandatory. A separate code of practice is planned for private companies. The government is also spending £11 million to triple the number of police vans fitted with live facial recognition.</p><p>Suppliers of facial recognition should prepare for independent bias testing as a condition of sale. Retailers and venues using the technology should expect their own code of practice to follow.</p><h3>Parts of Europe’s AI law now apply in Northern Ireland</h3><p>The British government <a href="https://www.gov.uk/government/publications/partial-application-of-the-eu-ai-act-in-northern-ireland">published guidance</a> on 4 November explaining that parts of the European Union’s AI Act apply in Northern Ireland from 10 November. This follows from the Windsor Framework, the post-Brexit agreement that keeps Northern Ireland aligned with some European product rules. The guidance is written for developers, manufacturers, importers and distributors, and helps them check whether their products are covered and which duties apply.</p><p>Any business selling products with built-in AI into Northern Ireland should read the guidance and check whether it is in scope. The rest of the United Kingdom is not affected.</p><p>That is this week’s ledger. The next issue arrives on Tuesday 17 November.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40d0ebef5686" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-ai-weekly-ledger/the-ai-weekly-ledger-40d0ebef5686">The AI Weekly Ledger</a> was originally published in <a href="https://medium.com/the-ai-weekly-ledger">The AI Weekly Ledger</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Google opens its computer-operating AI to every developer, as DeepSeek starts charging for search…]]></title>
            <link>https://medium.com/the-ai-weekly-ledger/google-opens-its-computer-operating-ai-to-every-developer-as-deepseek-starts-charging-for-search-71d30018a727?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/71d30018a727</guid>
            <category><![CDATA[deep]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 07:33:06 GMT</pubDate>
            <atom:updated>2026-11-06T07:33:06.075Z</atom:updated>
            <content:encoded><![CDATA[<h3>Google opens its computer-operating AI to every developer, as DeepSeek starts charging for search by AI tools.</h3><p>Google made its AI program that operates a computer on a user’s behalf available to all developers on 30 October, and DeepSeek began charging for its web search and page-fetch tools on 1 November. Both changes affect what it costs and what it takes to build software that acts on its own.</p><p>Run notes: this was an unattended run covering 27 October to 3 November 2026. Several pages I wanted (Axios, Real Python, the UK Society for Computers and Law round-up, the technology.org report, and UK government search pages) could not be opened, so I left out anything that depended only on them. The United States section is empty because I found nothing both new and verifiable. Items marked “reported” rest on a secondary outlet and I say what is unchecked.</p><h3>Research</h3><p><strong>Researchers show that web pages can trick AI assistants into believing a message came from the user</strong></p><p>A paper posted on 29 October by Ezra Edelman, Benjamin Edelman, Surbhi Goel and Eran Malach studies “role spoofing”. A hostile web page or tool result includes fake formatting and wording that make an AI model treat the text as an instruction from its user. The authors report that training a model against attacks based on wording stops those but leaves attacks based on fake structure working, and they propose a detector that reads the model’s internal signals to tell who is speaking. The abstract gives no success rates, and the paper is a preprint that has not been peer reviewed (<a href="https://arxiv.org/abs/2610.40396">arXiv</a>).</p><p>Anyone letting an AI assistant read the web or open documents should assume that text it reads can pose as the user, and should limit what the assistant can do without a human check.</p><p><strong>Microsoft and Arizona State University researchers propose a way to make AI models resist step-by-step jailbreak attempts</strong></p><p>A paper posted on 29 October, “ShallowDefend”, has the model first assess what the user is after, then record an explicit safe or unsafe decision, then reply. The decision step is trained with reinforcement learning, which rewards good outcomes. The authors say it gave the strongest protection against multi-turn attacks (conversations that coax a model across several messages) in their comparison across two model families and fourteen tests, with a modest loss in helpfulness. The abstract gives no figures, and the result is the authors’ own claim and unreplicated (<a href="https://arxiv.org/abs/2610.40627">arXiv</a>).</p><p>If you buy or build a chat product, ask the supplier how it was tested against multi-turn attacks, since single-message tests understate the risk.</p><p>I found no third item that was both new and checked against a primary source.</p><h3>Products and funding</h3><p><strong>Google makes AI that operates a computer generally available to developers</strong></p><p>Google’s developer release notes say that on 30 October the “Computer Use” feature, which lets a Gemini model look at a screen and click and type to complete a task, moved out of preview and into general availability in both of its programming interfaces. The notes do not state pricing or usage limits for it (<a href="https://ai.google.dev/gemini-api/docs/changelog">Google release notes</a>).</p><p>Operators who piloted computer-driving agents can now put them in production with a support commitment from Google. Check the price and the error rate on your own tasks before you commit.</p><p><strong>DeepSeek starts charging for web search and page fetching by its AI tools</strong></p><p>DeepSeek’s pricing page lists $0.75 per 1,000 successful search requests and the same for page fetches, with billing per request from 00:00 Beijing time on 1 November 2026. The page says there is no charge when the tools are used through DeepSeek’s own agent software, Claude Code or Codex. The page has no revision date, so I cannot say when that wording was written. It also shows that token prices are halved outside weekday peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC (<a href="https://api-docs.deepseek.com/quick_start/pricing/">DeepSeek pricing</a>).</p><p>Teams running agents on DeepSeek through their own software should budget for per-request search costs now. Teams that can batch work into off-peak hours can halve token spend.</p><p><strong>Cohere reported to have raised about $5.1 billion at a roughly $26 billion valuation (reported)</strong></p><p>The Canadian company that sells AI models to businesses is said to have started a funding round led by Germany’s Schwarz Group, with the Government of Canada, Bell Ventures and Bullhound Capital as new investors, Nvidia and Inovia among existing ones. The only source I opened is a single secondary site, which links no Cohere announcement, says the round “targets up to” C$7 billion, and lists investors inconsistently between its summary and its text. It also says Cohere’s North product revenue grew 165 percent year on year, which I could not check (<a href="https://entarabi.com/en/2026/11/cohere-raises-5-1-billion-in-series-e-funding-at-a-26-billion-valuation/">entARABI</a>).</p><p>Treat the amount and valuation as unconfirmed until Cohere or a major wire service publishes them. If true, it is the largest sign this year of governments and buyers funding a non-American model supplier.</p><h3>Regulation and policy</h3><p><strong>British government reported to be taking powers over AI companies through a cyber security bill (reported)</strong></p><p>An opinion piece in UnHerd on 24 October says the government tabled a new clause in the Cyber Security and Resilience Bill that would let it inspect, test and sanction AI companies located anywhere, without saying how (<a href="https://unherd.com/newsroom/britain-is-quietly-seeking-powers-over-foreign-ai-firms/">UnHerd</a>). A 2 November article from the consultancy Resultsense says the bill was listed for its third reading in the House of Lords on 2 November, and that the author could not retrieve the current Lords text (<a href="https://resultsense.com/insights/2026-11-02-ai-governance-supply-chain-risk">Resultsense</a>). I could not confirm whether the stage took place or the final wording. The clause itself dates from before this week.</p><p>Firms selling AI into Britain should read the Lords text on the Parliament website before assuming either that they are in scope or that they are not.</p><p><strong>British competition regulator reported to have opened an investigation into Meta’s terms for business content (reported)</strong></p><p>An aggregator says the Competition and Markets Authority, which enforces competition law in the United Kingdom, announced on 30 October an investigation into whether Meta’s terms let it use businesses’ and publishers’ content for its products including AI features, and that no finding of wrongdoing has been made. I could not open the government page, so the scope and legal basis are unchecked (<a href="https://aidapted.ro/en/articles/ai-news-gia-deepseek-openai-november-1-2026/">AIdapted</a>).</p><p>Publishers and advertisers on Meta should look for the official notice and check whether the investigation is inviting evidence.</p><p><strong>European Union and United States: nothing new and verified this week</strong></p><p>The European Commission’s AI Office sent information requests to OpenAI and Anthropic after incidents in which AI agents got out of their test environments. That was reported on 14 October, outside this week, and rests on one outlet that cites no Commission document (<a href="https://www.euronews.com/2026/10/14/eus-ai-rules-face-first-enforcement-test-over-lab-incidents">Euronews</a>). I found no new European or American legislation, court ruling, enforcement action or export rule dated 27 October to 3 November that I could verify.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71d30018a727" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-ai-weekly-ledger/google-opens-its-computer-operating-ai-to-every-developer-as-deepseek-starts-charging-for-search-71d30018a727">Google opens its computer-operating AI to every developer, as DeepSeek starts charging for search…</a> was originally published in <a href="https://medium.com/the-ai-weekly-ledger">The AI Weekly Ledger</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The New Target for Marketers Isn’t You — It’s Your AI Agent]]></title>
            <link>https://timchosen.medium.com/the-new-target-for-marketers-isnt-you-it-s-your-ai-agent-09e505c006ba?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/09e505c006ba</guid>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 12:30:14 GMT</pubDate>
            <atom:updated>2026-10-23T12:30:14.799Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>In 2012 I Predicted People Would Pay for Less Information. Here’s What is Coming Next.</blockquote><p><strong>The attention economy was only the first phase. The next one is already here.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XgcycpHRbfZEjZyw" /></figure><p>In 2012 and 2013, I made a prediction that sounded rather counter-intutive at the time,</p><blockquote>the world was heading toward a place where people would pay for <strong>less</strong> information.</blockquote><p>The dominant internet businesses then were built on the opposite premise. People paid for access to content, to data, to platforms. The model then was simply more information, more value. This then created an incentive for content generation and with it, the falling cost of publishing. That then created a flood of noise, and I reached a different conclusion. When information becomes abundant to the point of being overwhelming, the scarce and valuable thing is no longer access. It is <strong>relevance</strong>. And people will eventually pay to protect their attention from the surplus.</p><p>That prediction has already played out in plain sight. Spam filters, ad blockers, app blockers, paid email clients that hide promotions, and browser extensions that strip out noise are all expressions of the same underlying shift. People did not just want better information. They started paying to see less of it.</p><p>What is less widely understood is that this was only the first phase. The next one is now underway, and it changes how information is filtered, how decisions are made, and who (or what) marketers and builders actually need to reach.</p><h4>My Past Attempts at the solution</h4><p><strong>Dizcovery Network and Proof of Relevance</strong></p><p>One of the reasons I started a project called Dizcovery Network in the 2016 was to explore what happens when you take that insight seriously and try to redesign the economics around it.</p><p>The core idea was straightforward on the surface and more radical underneath</p><blockquote><strong>Get paid to browse the web.</strong></blockquote><p>We used early machine learning to model what individual users actually cared about: their interests, their patterns of engagement, the kinds of information they found useful. When someone paid to promote content through the platform, the system would show that content to people who might find it relevant. If a user did not find the content useful, they were compensated from the money the promoter had paid. If they did find it useful, they were not paid, but the promoter received genuine engagement rather than forced exposure.</p><p>The system ran in the background. Users simply used the platform as normal. Saving, sharing, ignoring, or engaging with content became signals. Those signals fed back into how relevance was assessed and how value was distributed.</p><p>What we were attempting to build was an advertising system based on <strong>proof of relevance (PoR).</strong> An advertiser could say, in effect, “I have a budget and a piece of content. I want it shown to people who are likely to care.” The platform would test that claim at scale. If the content performed poorly with the audience it was shown to, the economics adjusted and users who had been subjected to irrelevant material were compensated. If it performed well, the promoter benefited from real attention and the platform learned.</p><p>Later, projects such as Basic Attention Token and the Brave browser brought related ideas about paying people for their attention into wider view. The underlying observation was the same: once information is effectively infinite, the value moves to whoever can reliably deliver relevance rather than simply volume.</p><p><strong>An Early Conversational Resume System — CVBOT.ML</strong></p><p>A few years later, around 2019, I built something that now looks more familiar than it did at the time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pbxbCk8FBC7cx-j-Dy7YAw.png" /><figcaption>Gemini Summary of my early CV Bot</figcaption></figure><p>I created a conversational system that helped people build a résumé or CV by chatting with it. The interaction was simple. A user would indicate they wanted to create a résumé. The system would ask for their name, location, work history, day-to-day responsibilities, results, education, and other standard details. The user answered in ordinary language. The system turned those answers into structured bullet points, coherent experience sections, and a clean document.</p><p>The more important design choice was that the output was optimised for Applicant Tracking Systems — the software most companies used (and still use) to automatically screen applications. At the time, if a résumé did not contain the expected keywords in the expected places, the ATS would often reject it before a human recruiter ever saw it. Presentation and actual qualifications were secondary to whether the automated filter could parse and score the document.</p><p>The system was built to translate natural conversation into language the ATS could read and rate highly. For many users, that difference determined whether they reached a human interviewer at all. This was years before the current generation of large language models and AI résumé tools. At the time it was practical automation applied to a high-stakes, rules-bound process. In retrospect, it was an early example of using conversational interfaces and structured output to navigate systems that humans find opaque.</p><h4>My Next Prediction — Interface Shift</h4><p>These earlier projects were attempts to work with the consequences of information abundance. The current shift is more fundamental, the interfaces through which people interact with information are being replaced by agents.</p><h4>Email clients are becoming less visible</h4><p>Traditional email clients are becoming less central for many people. AI agents can already triage inboxes, identify what requires attention, draft replies, and surface only the items that need human judgment. The human reviews a summary or a draft, makes edits or gives an instruction, and the agent handles the rest. Opening a full email client and scanning dozens of messages is starting to feel like an unnecessary intermediate step. For a growing number of people, the primary interface is no longer the inbox itself but the agent that manages it.</p><h4>Traditional search is moving</h4><p>Search is following a similar path. A large and increasing share of research, comparison, and decision support now happens inside AI systems rather than through traditional search engines. Even when people still use search, AI-generated summaries frequently appear first. The pattern is moving from “<strong><em>I search, scan results, and click links</em></strong>” to “<strong><em>I ask an agent and receive a recommendation or a shortlist</em></strong>.” In many cases, the human never sees the full set of options the agent considered.</p><h4>Marketing to Agents Instead of Humans</h4><blockquote>This has a direct consequence for anyone building products, creating content, or running marketing.</blockquote><p>A growing portion of the research and shortlisting that used to happen in the open is now happening inside agent workflows. People ask an agent to compare options, summarise trade-offs, or recommend a path. The agent performs the work and returns a conclusion. The human often never sees the intermediate queries, the sources that were discarded, or the full reasoning.</p><p>As a result, the effective audience for a significant amount of product information, positioning, and content is no longer only the end user. It is also the agents that are filtering and recommending on the user’s behalf. If information is not clear, structured, and useful to those agents, it may never enter the consideration set that the human eventually sees.</p><p>This does not mean traditional marketing disappears. It means the distribution and evaluation layer is changing. SEO as it was practiced for human searchers is less decisive when the queries and the comparisons are happening inside systems that do not surface every result the same way. The practical question becomes whether an agent can reliably extract accurate, relevant information about what you offer and decide that it is worth including in its recommendation.</p><p><strong>New Risks When Agents Sit in the Decision Path</strong></p><p>The same shift creates new categories of risk, particularly around processes that involve money, approvals, or sensitive actions. A good example is <strong>Business email compromise (BCE) </strong>has historically relied on social engineering directed at humans such as impersonating a vendor or executive to trick someone into changing payment details.</p><p>When agents handle more of the triage, drafting, and initial evaluation, the attack surface moves with them. An attacker no longer needs to fully convince the human. They need to influence the agent’s context, the documents it reads, or the instructions it receives so that it surfaces a fraudulent change as legitimate. The human then reviews a summary or a pre-drafted message and approves it.</p><p>Similar issues can arise with refunds, data sharing, or other actions an agent is authorised to initiate or recommend. If systems are not designed with the assumption that agents can be given misleading context, coerced through carefully crafted inputs, or simply mistaken, the efficiency gains come with new exposure.</p><p>These are not entirely novel problems. They are familiar patterns of deception and process weakness adapted to a layer that sits between the human and the final action. The difference is that the human is less likely to see the full sequence of events.</p><h4><strong>Conclusion</strong></h4><p>In the early 2010s, the prediction was that people would begin paying to reduce their exposure to information. That has happened through blockers, filters, and paid tools that prioritise signal over volume. Early projects explored ways to price relevance more directly and to use conversational interfaces to navigate opaque systems.</p><p>The current phase hands more of the filtering, research, and initial decision-making to agents. Humans increasingly receive summaries and make final approvals rather than performing the intermediate work themselves. This changes the practical requirements for how information is structured, how products are positioned, and how processes that involve trust or money are secured.</p><p>The abundance of information did not make relevance less important. It made the systems that determine relevance more powerful. Those systems are now increasingly automated, and the humans who ultimately act on their outputs are one step further removed from the raw inputs. Understanding that distance, and designing for it is becoming a practical necessity rather than a theoretical concern.</p><p>This piece is based on ideas and projects I worked on in the 2010s and observations about how those dynamics are evolving with current AI systems. I plan to expand on the practical side such as structuring information for agents and reducing the associated risks in follow-up writing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=09e505c006ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[There Is No Intelligence in Artificial Intelligence]]></title>
            <link>https://timchosen.medium.com/there-is-no-intelligence-in-artificial-intelligence-267ce742901f?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/267ce742901f</guid>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 14:16:20 GMT</pubDate>
            <atom:updated>2026-10-12T14:16:20.940Z</atom:updated>
            <content:encoded><![CDATA[<p>A few days ago, I read a LinkedIn post that made me stop. As I remember it, a founder and CEO said he had told engineers to stop reviewing code produced by AI agents. He felt the tools had become good enough that human review was now slowing the work down. The project, he said, was in fintech.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3_JCO7tl990dcrW5" /><figcaption>Photo by <a href="https://unsplash.com/@cbpsc1?utm_source=medium&amp;utm_medium=referral">Clint Patterson</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>What concerned me was the reasoning. Being impressed by the code an AI produces is not enough to establish that it is ready to handle money or sensitive customer information. I do not know what other checks that team had in place, so I cannot judge the whole project from a post. But removing review needs a stronger justification than confidence in the tool.</p><p>That brought me back to a question I think we need to ask more often: what do we actually mean when we call these systems intelligent?</p><p>This was not meant to be a clickbait, but I agree that title needs an explanation. I’m not saying AI cannot solve problems or do useful work. I’m questioning the assumption that an answer which sounds intelligent must come from the same kind of understanding we expect from a person.</p><p>My point is practical. If we understand how these systems produce answers, we can give them better instructions, provide the right information and reduce some of the errors we call hallucinations. We can also get better at spotting when an answer needs checking and making informed decisions about the output they produce.</p><p>I’m talking here about the language models behind AI chat tools. Artificial intelligence is a much broader field, and a description of how a chatbot writes does not explain every kind of AI.</p><p><strong>Start with Humpty Dumpty</strong></p><p>If I say, “Humpty Dumpty sat on a…” you will probably finish it with “wall”, assuming you know the rhyme.</p><p>Now imagine asking someone who has never heard it. They could be exceptionally intelligent and still suggest a chair, a rock or a boat. Missing the familiar answer would tell us very little about their intelligence. They simply don’t have that particular reference.</p><p>A language model trained on the familiar rhyme is likely to produce ‘wall’ too. In that setting, it is a very likely continuation. But if I ask for a new version set at sea, ‘boat’ might become a more suitable answer. The words around the request change what makes sense.</p><p>That is a small example of why context matters. The same opening can lead to different answers depending on what we ask the model to do.</p><p><strong>Prediction can produce a convincing answer</strong></p><p>A language model generates text in small pieces called tokens. A token might be a word, part of a word or punctuation. At each step, the model assigns <strong>probabilities</strong> to <strong>possible</strong> <strong>next tokens,</strong> based on its <strong>training</strong> and the <strong>context available </strong>to it. The system then selects one and continues.</p><p>This does not mean it is only copying sentences. Learning to predict language can also produce useful representations of facts, relationships and ways of solving problems. It can combine those patterns in situations it has not encountered in exactly the same form before.</p><p>But the probability of a piece of text coming next is not the probability that its claim is true. A sentence can fit the conversation beautifully and still contain a made-up fact. Fluent writing alone gives us no way to tell the difference.</p><p>That is where people can get caught out. We read an answer with a confident tone, clear explanations and perhaps a few references, and assume someone has done the checking. That assumption needs testing.</p><p><strong>Take the article about walking</strong></p><p>Suppose I ask an AI chat bot to: “Write an article about the importance of walking 10,000 steps a day.”</p><p>The model can produce a convincing article using patterns and information learned from writing about exercise and the human body. That does not mean it has just looked up the research, compared the findings or checked whether my chosen number is the right starting point.</p><p>My prompt has already pushed it towards a particular argument. I have asked it to explain the importance of the target, rather than examine the evidence for it. It might challenge that assumption, but I should not depend on it doing so.</p><p><strong>A better brief would be:</strong></p><p>“<em>Write a plain-English article for adults who spend much of their working day sitting. Examine the evidence about daily walking and whether 10,000 steps is a useful target. Use the research documents I attach. Separate the findings from your interpretation, explain important limitations and identify which document supports each main claim. If the documents do not answer something, say so. Do not invent studies or figures</em>.”</p><p>I would then need to attach the actual documents. If I wanted fresh research, I would use a tool that can search and ask it to find and check the sources first. Merely asking for ‘the latest studies’ does not give a model internet access.</p><p>The improved brief specifies the audience and turns an assumed conclusion into a question to investigate. It also gives me a way to check the answer. I can open a source and see whether it supports the claim.</p><p><strong>Give it the information your task depends on</strong></p><p>The same principle applies when building software. ‘<strong>Build me a booking system</strong>’ leaves a lot unanswered. The model has to make assumptions about users, appointments, payments and the tools already in use.</p><p>If I explain that each appointment needs both a staff member and a room, that two bookings cannot use either resource at the same time, and that the business takes a deposit, the task becomes much clearer. If I also provide the relevant code and requirements, it has something concrete to work with.</p><p>That can improve the fit of the answer. It still does not prove the code works. I need to test the booking rules, including what happens when two people try to reserve the same slot.</p><p>For a letter, the useful context could be the actual dates, what happened, who will read it and what outcome I want. Without those details, a model may fill the gaps with assumptions. Some will be harmless. Others could change the meaning of the letter.</p><p>More text is not automatically better context. Outdated documents, conflicting instructions or incorrect figures can make the result worse. Give the model the information that matters, identify the authoritative source and make missing details visible.</p><p><strong>Making use of the backend (Advanced) settings</strong></p><p>If you are a developer and have access to Anthropic, OpenAI or similar developer API tools, there are also settings that you can adjust, depending on the model. This settings can change how the output of the system.</p><p><strong>Temparature</strong>: Temperature changes how the system samples from possible next tokens. Lower values generally favour the more likely choices. Higher values give less likely alternatives more opportunity. That can affect variation, but it is not an intelligence or accuracy dial. A low-temperature answer can still be wrong.</p><p><strong>Top-p: </strong>Top-p limits the candidate pool according to cumulative probability. Think of it as deciding how much of the field of possible continuations stays in contention. It also changes selection, not whether the underlying facts have been checked.</p><p>Fine-tuning is a separate process. It involves additional training on examples to change learned behaviour. Providing documents in a prompt gives the model material to use for that task; it does not ordinarily retrain the model’s underlying weights.</p><p>For an everyday user, a clear brief and suitable reference material are a practical place to start. You do not need to adjust developer settings to improve those.</p><p><strong>Making the output easier to check</strong></p><p>Ask the model to identify supporting passages, distinguish evidence from assumptions and acknowledge gaps. Where appropriate, limit the answer to the supplied documents. These are among the approaches Anthropic recommends for reducing hallucinations, with the explicit warning that they do not eliminate them.</p><p>Treat references as things to inspect. A source can exist without supporting the claim attached to it. Asking the same model to check its answer may help catch mistakes, but its agreement with itself is not independent confirmation.</p><p>A weak prompt is not the explanation for every failure. Models can make mistakes even when the instructions and source material are excellent. Responsibility does not disappear just because we wrote a careful brief.</p><p>That is also why the LinkedIn post bothered me. Better context can help an agent write the code we need, but somebody still has to decide whether the requirements are complete and whether the checks are good enough. In a payment system, that includes what happens when a request is repeated, a connection fails or someone tries to access another customer’s records. The appropriate level of review depends on what could go wrong. Confidence in the generator cannot settle that question on its own.</p><p><strong>Conclusion<br></strong>What do I want people to take from this piece?</p><blockquote>There is useful intelligence in what these systems can do. We should be more careful about what we infer from the way they speak.</blockquote><p>Give them a clear task, relevant evidence and room to say something is missing. Then check the parts you intend to rely on.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=267ce742901f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Got Tired of Leaving My Editor to Test APIs, So I Built My Own Client]]></title>
            <link>https://timchosen.medium.com/i-got-tired-of-leaving-my-editor-to-test-apis-so-i-built-my-own-client-a2ddd05acf8a?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/a2ddd05acf8a</guid>
            <category><![CDATA[graphql]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[postman]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[rest-api]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Mon, 13 Apr 2026 12:42:47 GMT</pubDate>
            <atom:updated>2026-04-13T12:42:47.760Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*966OrPteuK2dhxZUJv-kaQ.png" /></figure><p>Every developer knows the rhythm. You’re deep in a codebase, flow state fully engaged, and you need to test an endpoint. So you alt-tab to Postman, wait for it to load, dig through your collections, set up the right environment, fire the request, copy the response, switch back to your editor, and try to remember what you were doing.</p><p>That context switch, small as it sounds, adds up. It breaks concentration. It fragments your workflow. And after years of living with it, I decided to build something different.</p><p><strong>Gud API</strong> is a REST and GraphQL client that runs entirely inside VS Code. No separate app. No browser tabs. No context switching. Just open a panel in your editor and start testing APIs right where you write code.</p><h3>The Itch That Started It All</h3><p>I’ve spent a lot of time building and consuming APIs. Like most developers, I relied on Postman for years. It’s a solid tool — nobody’s arguing that. But it kept growing into something bigger than what I needed. What started as a lightweight HTTP client became a full platform with collaboration features, cloud accounts, and a login screen before you can send a GET request.</p><p>I didn’t need a platform. I needed a fast, focused way to fire off requests and see what came back, without leaving the place where I was already working.</p><p>The idea was simple: what if your API client was just another panel in your editor, like the terminal or the file explorer? Something that understood your workflow because it <em>was</em> part of your workflow.</p><p>That’s how Gud API was born.</p><h3>What It Actually Does</h3><p>At its core, Gud API is a full-featured API client. It supports every HTTP method you’d expect — GET, POST, PUT, PATCH, DELETE, OPTIONS, HEAD — with a clean request builder that handles query parameters, headers, authentication, and multiple body types including JSON, form data, multipart uploads, and plain text.</p><p>But the parts I’m most proud of are the details that make day-to-day API work smoother.</p><p><strong>GraphQL is a first-class citizen.</strong> There’s a dedicated editor with separate Query and Variables panes. When you switch to GraphQL mode, the extension automatically sets the method to POST and the content type to application/json. No fiddling with headers — just write your query and send.</p><p><strong>The response viewer does more than dump JSON.</strong> You get status codes, response times, and payload sizes at a glance. Toggle between pretty-printed, raw, and HTML preview modes. Search through response bodies with match counts and keyboard navigation. There’s even a cookies tab that shows exactly what the server sent back.</p><p><strong>Code generation turns any request into working code.</strong> One click gives you a ready-to-paste snippet in cURL, JavaScript (fetch), Python (requests), or Node.js (axios). It’s a small feature that saves a surprising amount of time when you’re writing integration code or documenting an API for a teammate.</p><h3>Collections, History, and Environments</h3><p>If you’ve used Postman or Thunder Client, the mental model is familiar. Save requests into collections, organize them in nested folders, and load any saved request back into the builder with a click.</p><p>Every request is also auto-saved to history — the last 100 entries — so even if you forget to save something, it’s still there in the sidebar with the method, URL, status code, and timestamp.</p><p>Environments work with a {{variable}} syntax that resolves automatically when you send a request. Define a base_url variable, set it to localhost:3000 in your dev environment and api.production.com in prod, and switch between them from a dropdown in the toolbar. Collection-level variables are supported too, with environment values taking priority when there&#39;s overlap.</p><h3>Switching From Other Tools Is Easy</h3><p>One thing I wanted to get right early was import support. If you’ve already invested time building collections in another tool, you shouldn’t have to start over.</p><p>Gud API can import Postman v2.x collections (including nested folders, headers, body, and collection variables), Thunder Client collections, and OpenAPI 3.x / Swagger 2.0 specs. With OpenAPI imports, paths become requests and tags become folders, so you get a usable collection structure immediately.</p><p>The latest release also added cURL import — paste a cURL command directly into the URL bar and it auto-fills everything: method, URL, headers, body, and auth. It’s the fastest way to go from a command you found in documentation to a live, editable request.</p><h3>It Looks Like Your Editor Because It Is Your Editor</h3><p>Gud API matches your VS Code theme automatically. If you use a dark theme, the client is dark. Light theme, high contrast — it adapts. There’s no jarring visual switch when you move between your code and your API panel.</p><p>It also works without a workspace open, storing data in VS Code’s global storage. That means your collections, history, cookies, and environments persist across sessions and projects.</p><h3>Free, Pro, and Team</h3><p>The free tier has no limits on requests, collections, folders, GraphQL, code generation, imports, history, or cookie management. You get one environment, which is enough for many workflows.</p><p>Pro unlocks unlimited environments and cloud sync across devices. Team adds shared collections with per-collection permissions, centralized billing, and invite codes so team members can join without individual licenses.</p><p>I wanted the free version to be genuinely useful — not a demo with a paywall behind every other feature. If you’re a solo developer testing APIs during local development, the free tier covers everything you need.</p><h3>What’s New in 0.5.0</h3><p>The most recent release added a few things I’m excited about. cURL import and export make it trivial to move requests between Gud API, documentation, and terminal workflows. There’s also a request timing sparkline in the sidebar — a tiny chart next to each saved request that visualizes your last 10 response times with color-coded indicators. It’s a quick way to spot if an endpoint is getting slower without running a formal benchmark.</p><h3>Getting Started</h3><p>Install Gud API from the <a href="https://marketplace.visualstudio.com/items?itemName=gudlab.gud-api">VS Code Marketplace</a> or <a href="https://open-vsx.org/extension/gudlab/gud-api">OpenVSX</a>. Click the Gud API icon in the activity bar, enter a URL, and press Send. That’s it.</p><p>If you have existing Postman, Thunder Client, or OpenAPI collections, click Import in the sidebar and point it at your JSON file. Your requests, folders, and variables come along for the ride.</p><h3>Why Build Another API Client?</h3><p>Fair question. The world doesn’t lack API tools. But most of them are standalone applications that exist outside the environment where you actually work. The ones that do live inside the editor tend to be limited — good for quick one-off requests but missing the collection management, environment handling, and import support that make a tool viable for daily use.</p><p>Gud API tries to sit in the middle: full-featured enough to replace a standalone client, but native to the editor so it doesn’t interrupt your flow.</p><p>If that sounds useful to you, give it a try. It’s free, it’s fast, and it’s already right where you need it.</p><p><em>Gud API is available on the </em><a href="https://marketplace.visualstudio.com/items?itemName=gudlab.gud-api"><em>VS Code Marketplace</em></a><em> and </em><a href="https://open-vsx.org/extension/gudlab/gud-api"><em>OpenVSX</em></a><em>. Learn more at </em><a href="https://gudlab.org/gudapi.html"><em>gudlab.org/gudapi</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a2ddd05acf8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Stopped Paying for Database GUIs. Here’s What I Use Instead.]]></title>
            <link>https://timchosen.medium.com/i-stopped-paying-for-database-guis-heres-what-i-use-instead-6b456c6f0167?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b456c6f0167</guid>
            <category><![CDATA[cursor-ide]]></category>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[database]]></category>
            <category><![CDATA[visual-studio-code]]></category>
            <category><![CDATA[windsurf]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Mon, 13 Apr 2026 07:28:41 GMT</pubDate>
            <atom:updated>2026-04-13T07:28:41.449Z</atom:updated>
            <content:encoded><![CDATA[<p><em>By Tim Chosen | @timchosen</em></p><p>I used to alt-tab out of Cursor dozens of times a day. Open pgAdmin to check a table. Switch to MongoDB Compass to peek at a document. Fire up DBeaver to run a quick query. Then context-switch back to my editor, losing my train of thought every single time.</p><p>One day I decided to count how many times I left my editor just to look at data. It was 37 times in a single workday. Thirty-seven interruptions. Thirty-seven broken flow states. So I did something about it.</p><h3>Meet GudDB</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Idu25rm0Z5DZvCBp8f_3YA.png" /></figure><p>GudDB is a free, lightweight database client that lives inside your favorite editor (VS Code, Cursor, Windsurf, Google Antigravity, etc). No separate app. No Electron overhead. No subscription. Just your databases, right where you write code.</p><p>It supports PostgreSQL, MySQL, MariaDB, SQL Server, SQLite, MongoDB, ClickHouse, ElasticSearch, and CockroachDB — all from a single extension that weighs under 500KB.</p><p>Yeah, you read that right. Under 500KB. DBeaver’s installer is over 100MB. TablePlus starts at $89. DataGrip is $229/year. GudDB is free and lighter than most of your node_modules dependencies.</p><h3>Why Another Database Extension?</h3><p>Fair question. The VS Code and Open VSX marketplace has a handful of database extensions already. Here’s why I wasn’t satisfied with any of them:</p><p>Most free options either support only one or two databases, lock core features behind a paywall after a trial period, or feel clunky — like someone wrapped a desktop app in a webview and called it a day.</p><p>I wanted something that felt native to VS Code. Something fast. Something that just worked when I needed to peek at a table or run a quick query without leaving my workflow.</p><h3>What You Actually Get</h3><p><strong>A real database explorer.</strong> Connect once, and your sidebar shows a navigable tree of connections, databases, schemas, tables, views, and columns. It feels like your file explorer, but for data.</p><p><strong>Inline editing that makes sense.</strong> Double-click a cell, change the value, done. Insert rows. Duplicate rows. Delete rows. No SQL required for basic CRUD — but the SQL editor is right there when you need it.</p><p><strong>A query editor with IntelliSense.</strong> Autocomplete for SQL keywords, table names, and column names. Hit Ctrl+Enter and your results show up in a clean grid below. It&#39;s not a toy — you can sort, filter, paginate, and work with large result sets comfortably.</p><p><strong>SSH tunneling built in.</strong> Need to connect to a production database behind a bastion host? Every server-based database in GudDB supports SSH tunnels out of the box. No extra config files, no third-party proxy tools.</p><p><strong>Export anything.</strong> Current page, all rows, selected rows, or a custom limit — export to CSV, JSON, or SQL INSERT statements. I use the SQL export constantly when moving test data between environments.</p><h3>The Feature That Made Me Grin: AI-Powered SQL</h3><p>This is the one that catches people off guard.</p><p>Hit Ctrl+Shift+G (or Cmd+Shift+G on Mac) and a prompt appears. Type what you want in plain English: &quot;Show me all users who signed up in the last 30 days and haven&#39;t made a purchase.&quot; GudDB generates the SQL for you.</p><p>It works with your Code Editor’s Copilot or Anthropic’s Claude — your choice. For developers who think in product terms rather than SQL syntax, this is a game changer. For SQL veterans, it’s a shortcut for those annoying multi-join queries you don’t want to write from scratch at 4pm on a Friday.</p><h3>How to Get Started</h3><p>Install it in under 10 seconds:</p><ol><li>Open VS Code (or Cursor, Windsurf, VSCodium — it works across editors)</li><li>Go to Extensions (Ctrl+Shift+X)</li><li>Search “GudDB”</li><li>Click Install</li></ol><p>That’s it. Click the database icon in your activity bar, add a connection, and you’re in. URI strings and JDBC URLs are auto-detected, so if you paste a connection string, it just figures it out.</p><h3>Who Is This For?</h3><p><strong>Full-stack developers</strong> who work with databases daily but don’t want to leave their editor. <strong>Indie hackers</strong> who don’t want to pay $200+/year for a database GUI. <strong>Backend engineers</strong> who need SSH tunnel support without the hassle. <strong>Anyone</strong> who’s tired of context-switching between their code and a separate database app.</p><h3>It’s Free. Actually Free.</h3><p>No freemium trap. No “free for 14 days.” No feature gates. GudDB is free to use, period. If you find it valuable and want to support continued development, there’s a Buy Me a Coffee link — but that’s entirely optional and doesn’t unlock anything extra.</p><h3>Try It, Break It, Tell Me What’s Missing</h3><p>GudDB is at version 0.4.4, and I’m actively developing it. I want to hear what works, what doesn’t, and what you wish it did. Open an issue on GitHub, leave a review on the VS Code marketplace, or just tweet at me.</p><p>The goal is simple: make the best free database client for VS Code. Not the most bloated. Not the most expensive. The best.</p><p><strong>Install GudDB:</strong> <a href="https://marketplace.visualstudio.com/items?itemName=gudlab.guddb">VS Code Marketplace</a> <strong>Report Issues:</strong> <a href="https://github.com/gudlab/guddb-database-client/issues">GitHub</a> <strong>Support Development:</strong> <a href="https://buymeacoffee.com/timchosen">Buy Me a Coffee</a></p><h3>Calling All Creators: Free License in Exchange for a Video</h3><p>Are you a YouTube creator, tech blogger, or content creator who covers developer tools? I’d love to partner with you.</p><p><strong>Make a video about GudDB — a walkthrough, review, tutorial, or comparison — and I’ll give you a free lifetime license</strong> to any future paid tier of GudDB when it launches. No strings attached. You get early access to premium features before anyone else, plus a permanent spot on our community wall of contributors.</p><p>Whether you have 500 subscribers or 500,000, I’m interested. The only requirement is that the video is published publicly and gives an honest take — positive or critical, I want real feedback.</p><p><strong>Interested?</strong> DM me on X/Twitter (<a href="https://x.com/timchosen">@timchosen</a>) or email me at <a href="/service/mailto:timchosen@gmail.com">timchosen@gmail.com</a> with a link to your channel. Let’s make something cool together.</p><p><em>If this was useful, give it a clap and follow me for more on developer tools, open source, and building in public. I’m working on a set of developer tools under the </em><a href="http://gudlab.org/"><em>GudLab</em></a><em> umbrella and sharing the journey. Please check more extensions and products for free on the </em><a href="http://gudlab.org/"><em>website</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b456c6f0167" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Recording Your Screen: How LaunchReel Turns Any Website Into Launch Videos Automatically]]></title>
            <link>https://timchosen.medium.com/stop-recording-your-screen-how-launchreel-turns-any-website-into-launch-videos-automatically-3536d3463588?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/3536d3463588</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 09:05:29 GMT</pubDate>
            <atom:updated>2026-03-16T09:05:29.231Z</atom:updated>
            <content:encoded><![CDATA[<p><em>By Tim Chosen | @timchosen</em></p><p>Every developer has been there. You’ve spent weeks building something incredible, and now you need a demo video. So you open your screen recorder, rehearse your clicks, stumble through three takes, edit out the awkward pauses, and end up with something that looks like it was made in 2012.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*SNJAvOFqDTYoAzgn9uAmhQ.gif" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FlSOHXzP3Vk8%3Fstart%3D1%26feature%3Doembed%26start%3D1&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DlSOHXzP3Vk8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FlSOHXzP3Vk8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bdf8d28ca6d9dd6c91543327a42c2896/href">https://medium.com/media/bdf8d28ca6d9dd6c91543327a42c2896/href</a></iframe><p>What if I told you there’s a better way?</p><h3>The Demo Video Problem</h3><p>When you’re launching a product — whether on Product Hunt, your landing page, or social media — the quality of your demo video often matters more than the quality of your code. Visitors decide within seconds whether your tool is worth trying, and a polished video can make or break that first impression.</p><p>But creating those videos is painful. Manual screen recording is inconsistent. You get cursor jitter, notification pop-ups, and that moment where you accidentally scroll past the feature you wanted to highlight. For indie developers and small teams, hiring a video editor just for a launch video is overkill.</p><p>That’s exactly why I built <strong>LaunchReel</strong>.</p><h3>What Is LaunchReel?</h3><p>LaunchReel is an open-source CLI tool that turns any website into professional launch videos and retina-quality screenshots — automatically. No screen recording, no video editing software, no re-takes.</p><p>You give it a URL. It studies your site, plans a series of demo scenarios, records them using real browser interactions, and exports polished MP4 videos ready for Product Hunt, Twitter, or your landing page.</p><p>It’s available as an npm package:</p><pre>npm install -g launchreel</pre><p>Or run it directly:</p><pre>npx launchreel study https://your-app.com</pre><h3>How It Works: Four Simple Stages</h3><p>LaunchReel follows a four-stage workflow that mirrors how a human would approach creating a demo video — but it does it consistently and reproducibly every time.</p><p><strong>1. Study</strong> — LaunchReel crawls your website using Puppeteer and Cheerio. It identifies pages, interactive elements, forms, and key features. Optionally, it can use an LLM (Claude, OpenAI, or Ollama) to understand what your product actually does and which features are worth showcasing.</p><p><strong>2. Plan</strong> — Based on what it learned, LaunchReel generates a launchreel.yaml configuration file. This is a human-readable, editable plan that describes each recording scenario: which pages to visit, what actions to take, and what to capture. You can tweak this file before recording.</p><p><strong>3. Record</strong> — Using Puppeteer for browser automation and rrweb for DOM session recording, LaunchReel executes each scenario. It navigates your site, clicks buttons, fills forms, and scrolls — all while capturing every interaction. It also takes retina-quality screenshots at key moments.</p><p><strong>4. Export</strong> — Finally, LaunchReel replays the recorded rrweb sessions in a headless browser, renders each frame, and stitches them into MP4 videos using ffmpeg. The result is a smooth, professional launch reel.</p><h3>The Killer Feature: Instruction-Driven Demos</h3><p>Here’s where LaunchReel gets really interesting. By default, the study command generates generic scenarios — scroll the page, hover over elements, take screenshots. Useful, but generic. The real power comes from the --instruction (or -i) flag, which lets you tell the AI exactly what workflow to demo in plain English.</p><p>Instead of hoping the crawler finds the right things to showcase, you describe the user journey you want to demonstrate:</p><pre># Demo a booking app&#39;s full workflow<br>launchreel study https://cal.com -i &quot;sign up, update availability for an event type, get the event link, book a meeting, accept the booking, show booking stats&quot;<br># Show the free trial signup experience<br>launchreel study https://myapp.com -i &quot;do a demo of the free trial signup flow&quot;<br># Walk through a specific feature<br>launchreel study https://myapp.com -i &quot;show how to setup the workspace feature and invite teammates&quot;</pre><p>The AI takes your instruction, cross-references it with the crawled site structure (pages, buttons, forms, navigation), and generates scenarios that click the right buttons, fill out forms with realistic data, navigate between pages, and walk through your workflow step by step. The result is a demo that looks like a real user interacting with your product — because it essentially is.</p><p>You still review the generated launchreel.yaml before recording, so you can adjust selectors or placeholder data. But the heavy lifting of figuring out which elements to interact with and in what order? That&#39;s handled for you.</p><h3>A Quick Example</h3><p>Let’s say you’ve built a project management tool and want to demo the onboarding flow:</p><pre># Step 1: Study your site with a specific instruction<br>launchreel study https://my-pm-tool.com -i &quot;create a new project, add three tasks, assign one to a teammate, and mark one as complete&quot;<br># Step 2: Review and edit the generated plan<br>cat launchreel.yaml<br># Step 3: Record the scenarios<br>launchreel record<br># Step 4: Export to MP4<br>launchreel export</pre><p>That’s it. Four commands. You’ll find your videos and screenshots in the output directory, ready to upload. And because you used --instruction, the video actually shows your product&#39;s value proposition — not just a generic scroll-through.</p><h3>Why Open Source?</h3><p>LaunchReel is part of GudLab, an open-source organization I founded. At GudLab, we believe developer tools should be accessible to everyone — not locked behind subscription paywalls. Our projects include GudForm (a free alternative to Typeform), GudCal (smart scheduling infrastructure), and GudDesk (customer support infrastructure for AI agents).</p><p>LaunchReel follows the same philosophy. It’s MIT-licensed, free to use, and open to contributions. If you’re launching a product and need a demo video, you shouldn’t have to pay hundreds of dollars for video editing software or spend hours doing manual screen captures.</p><h3>Who Is This For?</h3><p>LaunchReel is built for anyone who needs to showcase a web application visually:</p><ul><li><strong>Indie hackers</strong> preparing for a Product Hunt launch</li><li><strong>Open-source maintainers</strong> who want better README screenshots and demo GIFs</li><li><strong>Startup founders</strong> creating pitch deck materials</li><li><strong>Developer advocates</strong> producing tutorial content</li><li><strong>Agencies</strong> generating deliverables for client presentations</li></ul><h3>Get Involved</h3><p>LaunchReel is brand new and actively being developed. Here’s how you can get started:</p><ul><li><strong>Install it:</strong> npm install -g launchreel</li><li><strong>Star the repo:</strong> <a href="https://github.com/gudlab/launchreel">github.com/gudlab/launchreel</a></li><li><strong>Report issues:</strong> Contributions, bug reports, and feature requests are all welcome</li><li><strong>Join the conversation:</strong> Follow <a href="https://medium.com/@timchosen">@timchosen</a> for updates</li></ul><p>The days of awkward screen recordings are over. Let your code speak for itself automatically.</p><p><em>Tim Chosen is a tech leader and the founder of GudLab and sonal-ai.com. He builds open-source tools for developers and writes about the intersection of AI, automation, and product development.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3536d3463588" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Manually Recording Your Product Demos. There’s a Better Way.]]></title>
            <link>https://timchosen.medium.com/stop-manually-recording-your-product-demos-theres-a-better-way-b92afb838823?source=rss-a3aaf36b7919------2</link>
            <guid isPermaLink="false">https://medium.com/p/b92afb838823</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Tim Chosen]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 09:05:14 GMT</pubDate>
            <atom:updated>2026-03-16T09:05:14.151Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How I built an open-source tool that turns any website into polished launch videos and screenshots with a single command.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aiFh1Jb5jBv_0KYoUy7r-w.png" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FlSOHXzP3Vk8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DlSOHXzP3Vk8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FlSOHXzP3Vk8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7797c23b419475f4f540c6bdf0387fed/href">https://medium.com/media/7797c23b419475f4f540c6bdf0387fed/href</a></iframe><p>You’ve spent months building your product. The code is solid. The features are ready. Now comes the part nobody warns you about: making it <em>look</em> ready.</p><p>You need a Product Hunt launch video. Social media clips. App store screenshots. A hero demo for your landing page. And suddenly, you’re spending hours screen-recording, re-recording when you make a mistake, trimming in iMovie, and still ending up with something that looks like it was made during a lunch break.</p><p>I know this pain intimately. I run multiple products under the GudLab umbrella, and every time I ship an update or launch a new tool, the same bottleneck appears: creating visual marketing assets.</p><p>So I built <strong>LaunchReel </strong>and I’m open-sourcing it.</p><h3>What is LaunchReel?</h3><p>LaunchReel is a CLI tool that crawls your website, records browser sessions using rrweb, and exports polished MP4 videos and retina screenshots. Three commands. That’s it.</p><pre>npx launchreel study https://myapp.com<br>npx launchreel record<br>npx launchreel export</pre><p>Your output folder now contains named screenshots, per-page videos, and a combined launch reel, ready to upload to Product Hunt, X, LinkedIn, or your landing page.</p><h3>Why I Built This</h3><p>Every indie maker, startup founder, and developer advocate faces the same problem: your product demo workflow is stuck in 2015. You’re manually clicking through your app while a screen recorder runs in the background. You mess up, you start over. You forget to scroll slowly. The cookie banner pops up mid-recording.</p><p>The irony? We automate everything else. Our deployments, our tests, our infrastructure. But the thing that actually convinces people to <em>try</em> our product - the visual first impression - is still a manual, painful process.</p><p>LaunchReel treats demo creation the way we treat everything else in modern development: as code.</p><h3>How It Actually Works</h3><p>LaunchReel follows a four-step pipeline:</p><p><strong>Study</strong> — Point it at your URL and it crawls your site, following navigation links and sitemaps. It extracts page structure and, optionally, uses an LLM (Claude, GPT-4o, or a local Ollama model) to analyse your features and generate an intelligent recording plan.</p><p><strong>Plan: </strong>The output is a human-readable YAML config. You can review it, tweak it, and version-control it alongside your codebase. Want to change the scroll speed on the homepage demo? Edit one line.</p><p><strong>Record</strong>: Puppeteer navigates your site while rrweb captures every DOM mutation. Screenshots are taken at the moments you define. Cookie banners get dismissed automatically. CSS animations are forced to completion so nothing renders half-visible.</p><p><strong>Export: </strong>The rrweb events are replayed in headless Chrome, frames are captured, and ffmpeg encodes everything into clean MP4 videos. Multiple scenarios get stitched into a single launch reel.</p><h3>The Config-as-Code Approach</h3><p>This is where LaunchReel gets interesting. Your entire recording plan lives in a YAML file:</p><pre>scenarios:<br>  - id: &quot;homepage&quot;<br>    name: &quot;Homepage Tour&quot;<br>    type: &quot;screenshot+video&quot;<br>    page: &quot;/&quot;<br>    actions:<br>      - dismiss_cookies: true<br>      - finish_animations: true<br>      - wait: 500<br>      - screenshot: { name: &quot;01-hero&quot; }<br>      - scroll: { to: 600, duration: 1500 }<br>      - screenshot: { name: &quot;02-features&quot; }</pre><p>This means your demo assets are <strong>reproducible</strong>. Ship a new version? Re-run the same config and get updated screenshots and videos that match your latest UI. No more “oh wait, that screenshot is from three releases ago.”</p><p>It also means your demo workflow is <strong>collaborative</strong>. A designer can tweak timings. A marketer can add scenarios for specific features. It’s all just YAML.</p><h3>Authentication? Handled.</h3><p>One of the biggest headaches with automated browser tools is dealing with logged-in states. LaunchReel supports three auth strategies out of the box: credential-based login (with environment variable interpolation for security), cookie-based auth via a manual login helper, and token injection into localStorage or sessionStorage.</p><p>Need to demo your dashboard? Your admin panel? Your authenticated onboarding flow? It just works.</p><h3>AI-Powered Site Analysis</h3><p>The study command can optionally use an LLM to analyse your site and generate optimised recording scenarios. It supports Claude, OpenAI, and local models via Ollama , so you can run the entire pipeline without sending any data to external APIs if you prefer.</p><p>The AI examines your site’s structure and suggests which pages to record, what interactions to capture, and how to sequence them for maximum impact. Think of it as having a product marketing consultant look at your site and plan your demo, except it takes seconds instead of days.</p><h3>Who Is This For?</h3><p><strong>Indie makers</strong> launching on Product Hunt who need a demo video but don’t have a video team.</p><p><strong>Startup founders</strong> who need consistent, up-to-date screenshots for their pitch decks and app store listings.</p><p><strong>Developer advocates</strong> who want to generate tutorial screenshots and demo clips as part of their CI pipeline.</p><p><strong>Marketing teams</strong> who are tired of asking engineering to “re-record that demo because we changed the hero copy.”</p><p><strong>Anyone</strong> who has ever thought: “I just need a quick video of my app doing the thing, why is this so hard?”</p><h3>Get Started in 60 Seconds</h3><p>You don’t even need to install it globally:</p><pre>npx launchreel study https://your-site.com<br># Review and edit launchreel.yaml<br>npx launchreel record<br>npx launchreel export</pre><p>Your videos and screenshots will be waiting in launchreel-output/.</p><p>The full source is on GitHub: <a href="https://github.com/gudlab/launchreel">github.com/gudlab/launchreel</a></p><h3>What’s Next</h3><p>LaunchReel is MIT-licensed and under active development. I’m working on additional export formats, smoother transitions between scenarios, and deeper AI integration for auto-generating optimal recording plans from your site’s content.</p><p>If you’ve ever lost hours to manual screen recording, give LaunchReel a try. Star the repo if it saves you time. Open an issue if it doesn’t.</p><p>Your product deserves a launch reel that matches the effort you put into building it.</p><p><em>LaunchReel is part of the </em><a href="https://github.com/gudlab"><em>GudLab</em></a><em> family of open-source developer tools. Follow along for more tools that solve the small but painful problems in the startup workflow.</em></p><p><em>If you found this useful, give the repo a ⭐ on </em><a href="https://github.com/gudlab/launchreel"><em>GitHub</em></a><em> as it helps more developers discover it.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b92afb838823" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>