It's called andrej-karpathy-skills. Worth saying up front: it's not actually from Karpathy. Someone took his public rants about where LLMs fall over when they write code and turned them into a single CLAUDE.md you drop into any project. That's the whole repo. One file. 178k stars. The reason it's blowing up is the four rules it gives Claude: 1. Think before coding. Surface your assumptions first instead of charging in. 2. Simplicity first. No speculative features, no "while I'm here" extras. 3. Surgical changes. Touch only what the task needs, leave the rest alone. 4. Goal-driven execution. Give Claude the success criteria and let it loop until it hits them, instead of dictating every step. That last one is the bit most people sleep on. You stop micromanaging the steps and start describing what "done" looks like. Claude is genuinely good at looping toward a target. It's worse at following a 12-step recipe you wrote. Drop the file in, restart Claude Code, and you'll feel the difference on the next task. Repo here: https://github.com/multica-ai/andrej-karpathy-skills Have a read and tell me which of the four your projects break most often. Mine's number 2.

Claude Code's biggest weakness is memory. On a big codebase it greps through your files every time you ask it something, burns tokens doing it, and forgets the layout the second the session ends. This video walks through Graphify, a free open source tool (nearly 70,000 stars) that fixes that by turning your whole repo into a knowledge graph Claude queries instead of crawling files. I also run a real token test so you can see the actual saving, not the hyped number. Here's what's inside: Step 1: What it actually does Graphify reads your repo once and builds a map. Not just that file A connects to file B, but why. Claude reads the map instead of grepping. Fewer tokens, more accurate answers, and it remembers the structure. Step 2: The three passes Pass 1 is your code, parsed locally with tree-sitter (classes, functions, imports, call graphs), no LLM, free. Pass 2 transcribes any audio/video locally with Whisper. Pass 3 uses Claude sub agents to read docs, PDFs and images and slot them into the graph. Step 3: Why it's not graph RAG Graphify uses no embeddings. It's best for mapping a codebase and how it's wired. A graph RAG system is better for questioning a huge pile of unstructured PDFs. Think of Graphify as sitting between an Obsidian vault and a full RAG setup. Step 4: Installing it Easiest path in Claude Code: copy the Graphify GitHub URL, paste it into Claude, and ask it to install Graphify for you. It comes with a skill that tells Claude which command to use, so you don't have to memorise anything. Step 5: The token test I cloned Papermark (a real, sizable repo), built the graph, then asked the same question twice. With Graphify: ~54k tokens. Without: ~108k. Same correct answer, about half the cost. Not the 70x people throw around, but a real saving that pays for itself. Resources I use: - Graphify (GitHub): https://github.com/safishamsi/graphify Build the map once, and with the commit hook installed it keeps itself current for free. That's the memory people keep talking about.

Here's the problem it solves. On a big repo, Claude greps your files every time you ask it something. It opens them, reads them, works out how they connect, and burns a load of tokens doing it. Then the session ends and it forgets the lot. Graphify reads your repo once and builds a knowledge graph. Claude queries that map instead of crawling files. Fewer tokens, more accurate answers, and it remembers the structure. How to try it this week: 1. Install it Grab the repo (github.com/safishamsi/graphify), paste the link into Claude Code, and say "install Graphify for me." It reads the repo and sets itself up. There are manual steps in the readme but you don't need them. 2. Build the map Run /graphify . in the folder you're working in. It pulls out your nodes (the pieces), edges (the connections) and communities (clusters of related code). It also shows you the most connected parts of the repo, so you get a tour before you've typed a prompt. 3. Make Claude actually use it Use graphify query or graphify explain so Claude answers off the map instead of guessing. Run graphify claude install to make it always on. 4. Keep it current Run graphify hook install and it rebuilds after every commit. That rebuild is free, no API cost, so the map never goes stale. What you get: Claude stops re-reading your whole project on every question. Cheaper, faster, and it actually knows how things are wired. Take 20 minutes this week, point it at a repo you actually work in, and ask it the same question with and without the map. Watch the token count. What codebase would you point it at first?

Anthropic just put Fable 5 in everyone's hands, so now you've got three coding models to choose from inside Claude Code. Most people leave it on default and never think about it. That's money left on the table in both directions: overpaying on simple jobs, underpowering the hard ones. Here's the rule I use. First, how to switch. Type /model in Claude Code and pick from the list. You can change it mid-session whenever the task changes. That's the whole mechanic; it takes two seconds. Now when to use which: 1. Haiku for the cheap, repetitive stuff Renaming files, simple edits, running through a checklist, quick lookups. Don't burn a frontier model on work a small model nails. It's faster and almost free. 2. Opus 4.8 as your daily driver This is where most of your real work lives. Building features, debugging, writing, planning. It's the best balance of smart and affordable, and it'll handle 90% of what you throw at it. 3. Fable 5 for the long, messy, high-stakes tasks Big multi-step builds where the cheaper models drift halfway through. Gnarly bugs nobody can crack. The work where one good output is worth more than the token cost. It's twice the price of Opus 4.8, so reach for it on purpose, not by accident. The simple version: match the model to the difficulty of the task, not the other way round. One timing note. Fable 5 is free on the paid plans until June 22. So the next couple of weeks are the cheapest chance you'll get to test it on your own real work and feel where it actually pulls ahead before it starts costing credits. Try this today: next time you start a session, ask yourself "is this a Haiku job, an Opus job, or a Fable job?" before you type the first prompt. What are you mostly running on right now, and have you noticed a difference when you switch up for the hard stuff?