<h3 style="font-size: 22px; font-weight: normal;">The #2 trending repo on GitHub this week isn't code. It's a single markdown file.</h3>forrestchang/andrej-karpathy-skill <a target="_blank" rel="noopener noreferrer nofollow" href="http://CLAUDE.md" style="text-decoration: underline;">CLAUDE.md</a> file — picked up 7,319 stars in seven days. It's not a framework. It's not a tool. It's not a package. You can read the entire repo in under ten minutes.The whole content is a list of rules telling Claude Code to stop over-engineering.<ol class="editorOrderedList"><li class="editor_editorListItem__FKzSL">State your assumptions before you write code.</li><li class="editor_editorListItem__FKzSL">Implement only what was requested.</li><li class="editor_editorListItem__FKzSL">Touch only the lines directly required.</li><li class="editor_editorListItem__FKzSL">Turn the task into a verifiable goal, then loop until the goal is met.</li></ol>Derived, the description says, from Andrej Karpathy's observations on LLM coding pitfalls. The rules themselves are almost embarrassingly simple. <ul class="editorBulletList"><li class="editor_editorListItem__FKzSL">"Don't add features I didn't ask for." </li><li class="editor_editorListItem__FKzSL">"Don't refactor things I didn't ask you to touch." </li><li class="editor_editorListItem__FKzSL">"Say when you don't know."</li></ul>7,319 developers starred it in a week. Now look at what sat next to it on the chart.#1 NousResearch/hermes-agent — "The agent that grows with you." 38,426 stars. A heavyweight self-improving agent framework.#4 multica — "Open-source managed agents platform. Turn coding agents into real teammates." 6,846 stars.#6 Archon — "The first open-source harness builder for AI coding. Make AI coding deterministic and repeatable." 2,962 stars.Every single repo in the top 10 is about getting AI to do more. Generate more. Self-improve more. Coordinate more teammates. Orchestrate more agents. Compound more skills.And sandwiched between all of them, at #2, is a 200-line markdown file whose entire purpose is to get the AI to do less.That's the week's real signal.<img src="https://d20jzv1ghtxgkk.cloudfront.net/images_temp/1/image-1777866042891.png" width="90%" draggable="true" style="max-width: 100%; width: 100%; height: auto; cursor: pointer;">Six months ago, the meta on this chart was agent frameworks. The race was who could build the most autonomous AI coworker. Every launch promised more autonomy, more capability, more "set it and walk away." And then the codebases started to rot.Talk to anyone who's been shipping with Copilot or Claude Code for six months and you'll hear the same thing. Not that the AI can't code. The AI codes fine. The problem is the AI codes too much. Try-catches around console.logs. Forty-line utilities that reimplement .map.. Variables named userPreferencesuserPreferences that actually hold session tokens. "Defensive" code paths for scenarios that can't happen. Abstractions over three lines of logic. Helper functions for things that didn't need a helper.The Reddit audit post that trended last week ("I audited 6 months of PRs after my team went all-in on AI code generation") hit 1,657 upvotes for one reason: <h3 style="font-size: 22px; font-weight: normal;">Everyone was already feeling this. </h3><h3 style="font-size: 22px; font-weight: normal;"></h3>The AI isn't wrong. It's just overbuilt.So this week the #2 repo on GitHub is someone — derived from Karpathy's observations — saying it out loud. Don't overbuild. Ask what the user actually wanted. State your assumptions. Touch only what's required.A single file.7,319 stars.<h3 style="font-size: 22px; font-weight: normal;">There's a larger pattern underneath this, and it's the pattern we keep coming back to while building Bunzee.</h3><h3 style="font-size: 22px; font-weight: normal;"></h3>The bottleneck in building has moved. It used to be: can I ship this? Now it's: should I ship this? AI has solved "can." The scarcity has shifted to judgment — what to build, what to leave out, what features your idea doesn't need, what entire product ideas aren't actually worth pursuing.Karpathy-skills is that judgment, applied at the code level. "Stop adding things I didn't ask for."<img src="https://d20jzv1ghtxgkk.cloudfront.net/images_temp/1/image-1777866063771.png" width="90%" draggable="true" style="max-width: 100%; width: 100%; height: auto; cursor: pointer;">Bunzee is that same judgment, applied one level earlier — at the idea level. Before you write a single line, before you even pick a framework, you talk through your idea and we compare it against our vector DB of real competitor data. Not a fantasy market report. Actual products, pricing, positioning, user complaints. You get TAM/SAM/SOM estimates grounded in what's already out there, an ERRC framework showing where the space is crowded and where there's real white space, a PRD, and an HTML MVP prototype. Ten minutes. Conversational.The point isn't to slow you down. The point is the same point Karpathy's list is making about code: the most valuable thing you can do right now is not generate more output. It's generate less of the wrong output.In 2026, a single file of "don't overbuild" gets 7,319 stars in a week.That's the mood. Build accordingly.

GitHub's #2 trending repo wasn't code — it was a single markdown file

#trend #item #business #github #repo #forrestchang/andrej-karpathy-skills #doless #AI #needs #market

My idea works in the market?