Andrej Karpathy joined Anthropic’s pre-training team this month. Around the same time, a community file distilling his coding-agent observations into a single CLAUDE.md crossed 60,000 stars. Four principles, no ceremony. So I opened my own CLAUDE.md, the one I’d built up over a year of corrections, and lined them up side by side.

Personal update: I’ve joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative.

— Andrej Karpathy, May 2026

I’d already written two of the four. I’d never written the other two. The gap turned out to be the whole point.

The Four Principles

The viral file isn’t Karpathy’s. It’s a community distillation of his observations, which is part of why it’s interesting: it’s what a crowd decided was load-bearing.

  • Think before coding. State assumptions, surface ambiguity, push back before typing.
  • Simplicity first. Minimum code for the stated problem. No speculative abstractions.
  • Surgical changes. Don’t touch adjacent code. Match the existing style. Remove only what your change orphaned.
  • Goal-driven execution. Turn “add validation” into “write failing tests, then loop until they pass.”

Where Mine Already Matched

Two of these I could have quoted from my own file before reading theirs.

My YAGNI section is simplicity first almost verbatim: only what’s asked, no unrequested features, fallbacks, or future-proof abstractions. My “no backwards-compatibility hacks” rule is surgical changes down to the specifics, including the oddly exact instruction to remove only the imports and variables your own edit orphaned, and to delete dead code rather than comment it out.

I didn’t copy these. Neither did whoever wrote the viral file. We both arrived at them because they’re what survives contact with a coding agent that will happily over-engineer and “improve” code you never asked it to touch. This is the same lesson Boris’s vanilla-but-intentional setup teaches: the fundamentals compound, and everyone serious converges on the same small set.

The rules you can recite from memory are the ones you’ve already internalised. The rules missing from your config are the shape of your blind spots.

The Two I’d Never Written Down

Think before coding and goal-driven execution were nowhere in my file. Not because I disagree, but because I do them by reflex and never bothered to encode them. That’s exactly why they were missing, and exactly why they matter.

The two I had written are defensive: they stop the agent doing damage. The two I’d left out are generative: they’re what make the agent useful unsupervised. “Think first” is Plan Mode as a habit. “Goal-driven” is the difference between an agent that does what you said and one that loops until the thing actually works. The leverage was in the principles I’d internalised so deeply I forgot to write them for the model, which can’t read my reflexes.

Why Only Four Survive

There’s a reason these distil to four and not forty. As I argued in your AGENTS.md is a liability, instruction compliance degrades as the file grows, and early instructions win. A CLAUDE.md is not a wishlist; it’s a context budget. The principles that survive aggressive pruning are the load-bearing ones, which is precisely the filter both the viral file and my own went through. Convergence isn’t a coincidence. It’s two pruning processes landing on the same skeleton.

What a CLAUDE.md Can’t Do

The honest limits, because a file full of good principles is not a substitute for judgement.

  • Rules aren’t guarantees. The model’s compliance is probabilistic. “Think before coding” in your config raises the odds; it does not enforce them.
  • Copying the viral file wholesale is cargo-culting. The value was in the distillation, not the artifact. A file you didn’t earn through your own corrections won’t fit your work, and it adds to the context budget regardless.
  • The principles are generic on purpose. They’re a floor, not a strategy. Your domain rules, the ones nobody else can write, are where the real configuration lives.
Audit by recall, not by copying

Don’t paste someone else’s CLAUDE.md. Do the reverse: read theirs, then try to recite yours from memory. The principles you can’t recall are the ones you haven’t internalised. The ones you’ve internalised but never wrote down are the ones your agent is missing. That gap is the only part worth copying.

Closing

The interesting thing about the viral file wasn’t that it was right. It was that I’d independently written half of it, which means the half I hadn’t was a measurement of my own blind spots. Karpathy going in-house at Anthropic will change the models. The more useful artifact this month was a 65-line file that gave a lot of us a mirror.

Related Posts