A single CLAUDE.md file has 201,979 stars, and it does not add a feature to Claude Code. It removes bad habits. https://github.com/multica-ai/andrej-karpathy-skills It comes from a post Andrej Karpathy wrote about where LLMs fall down when they write code. His complaint was not that they are dumb. It was that they make a wrong assumption on your behalf, never check it, and then confidently build on top of it. And that they write 1,000 lines where 100 would do. Somebody turned that into four rules and shipped it as a plugin. Think before coding. State your assumptions out loud, ask instead of guessing, push back if there is a simpler way. Simplicity first. No features nobody asked for, no abstraction for something used once, no error handling for things that cannot happen. Surgical changes. Touch only what the task needs. Do not tidy up the code next to it, do not delete dead code you happened to notice. Goal-driven execution. Stop giving instructions and start giving success criteria. Instead of fix the bug, say write a test that reproduces it then make it pass. Karpathy's line is that models are exceptionally good at looping until they hit a specific goal, so give them one. Two ways in. As a plugin, run /plugin marketplace add forrestchang/andrej-karpathy-skills then /plugin install andrej-karpathy-skills@karpathy-skills, and it applies everywhere. Or drop the CLAUDE.md straight into one project. Straight with you: I have not run this on a real build yet, and the repo has not been updated since April. So take it as an idea rather than a recommendation. The idea is worth having even if you never install it. Most people write prompts as instructions. The fourth principle says write them as a finish line instead, then let the model run at it. That one change costs nothing and you can test it on your next prompt. Try it and tell me if your outputs get shorter. That is the tell.

It is a design language for AI coding agents. Install it once, run one setup command, and the frontends your agent builds stop looking like every other AI frontend. Free, Apache licensed, 57,000 stars. Works in Claude Code, Cursor and Codex. Here is the problem it solves, and it is worth understanding even if you never install it. Every model was trained on the same SaaS templates. So skip the design guidance and you get the same handful of tells on every project. Inter for every heading and every body line. Purple to blue gradients. Cards nested inside cards. Grey text sitting on a coloured background. A rounded square icon tile above every heading. Once you can name those five you cannot unsee them. You will spot them on half the sites you open this week. Setup is two commands. Run npx impeccable install from your project root, then /impeccable init inside your coding tool. Init asks whether the surface is brand or product, then writes PRODUCT.md and DESIGN.md into your repo. Those two files are the actual trick. Every later command reads them, so your agent knows the audience, the voice, the colours and the anti-references without you re-explaining it on every prompt. After that you get 23 commands as shared vocabulary with your AI. polish, audit, critique, distill, animate, bolder, quieter. Instead of writing a paragraph about what feels off, you say /impeccable quieter and it knows what you mean. The part I did not expect: 59 of the checks are deterministic rules rather than model calls. The CLI and the browser extension run them with no LLM and no API key. So the audit costs nothing and gives you the same answer every time. Repo: https://github.com/pbakaus/impeccable Docs: https://impeccable.style The takeaway is not the tool though. It is that your agent produces generic design because nothing in your repo tells it who the thing is for. Write that down once and the output changes on every prompt after it.

One product photo goes in. Nine on-brand ads come out, then a video ad, then a UGC ad with a creator talking to camera. All from one skill, inside Claude Code. I built Ad Factory OS because I was tired of juggling a separate skill for images, another for video, another for UGC. This is the one stop shop. It is free and it is in the classroom here. If you have a product and one decent photo of it, you have everything you need to run this today. send this prompt to Claude Code **copy below this line** I'm giving you a skill called Ad Factory OS. Get the files with: git clone https://github.com/automatorsplus/ad-factory-os-skill Follow the README to set it up. Recommend how it would best apply to our setup in plain language, and ask me a few questions to clarify my intent before you install it. Drop your ads below... 🚀

It makes your coding agent answer you in grunts. The README says that cuts 65% of output tokens. That number is real. It is also not the number most of you would get. The repo is the one that tells you so. Here it is: https://github.com/JuliusBrussee/caveman Here is what it does. One command, works on Claude Code and about thirty other agents. Before, your agent answers like this: Sure! I'd be happy to help you with that. The issue you're experiencing is most likely caused by your authentication middleware not properly validating the token expiry. Let me take a look and suggest a fix. After, it answers like this: Bug in auth middleware. Token expiry check use < not <=. Fix: Code, commands and error strings come through byte for byte identical. Only the waffle gets squeezed. The 65% comes from ten chat-style questions. Ask it to explain a React re-render bug and the answer drops from 1,180 tokens to 159. Then JetBrains ran it properly. 86 real coding tasks, auto-graded, Claude Code on Sonnet 5. The saving was 8.5%. Both numbers are honest. They measured different work. And the gap is the whole lesson. When an agent explains something to you, the prose IS the answer, so squeezing the prose squeezes everything. When an agent works through a repo on its own, the prose is a thin strip between tool calls. The tool calls, diffs and code are untouchable by design. There is barely anything left to cut. There is a bigger version of this. Your agentic bill is mostly input tokens, not output. A skill that only shrinks output cannot touch the side you are actually paying for. The 65% was never going to show up on your invoice. The bit worth stealing has nothing to do with installing it. Next time a tool quotes you a number, ask what workload it was measured on, then ask whether that is the workload you have. Chat-style prose and unattended repo work are different jobs. A benchmark run on one tells you very little about the other.

It is the distribution layer. You write the content wherever you write it. Blotato pushes it live to every platform on a schedule, from one place. Not clever on its own. What makes it worth an hour of your week is that it runs as an MCP server, so Claude does the posting instead of you clicking round a dashboard. I set it up with a client back in June. By the last session she was scheduling a week of content by talking to Claude. No tabs, no copy and paste, no logging into five apps to put the same thing out four times. Four things I would tell you before you start, because all four cost me time. Always pull the account list before you post. Every connected account has an ID. That ID is what the post attaches to. Guess it or hardcode it and your post goes to the wrong place or nowhere at all. TikTok wants seven separate true or false flags on every single post. Privacy level, comments, duet, stitch, branded content, your brand, AI generated. Miss one and it rejects the lot. Every platform has its own version of this, so read what the account list tells you rather than assuming.