The instruction files coding agents read at the start of every session, measured across public GitHub repos. Read the full write-up →
- The median CLAUDE.md is about 1,379 tokens, loaded every session.
- Files that kept the
/initopening spend far more of that on overviews. - Paste yours to see where it sits. It runs in your browser.
Try this: replace the sample with your file and read the cards.
Check your own file
This page needs JavaScript.
How big are they?
Share of repos whose file is a given size or smaller. Popular repos (10+ stars) and small, personal ones (0 to 9 stars) are drawn separately. The median CLAUDE.md is tokens and the 90th percentile is . The median AGENTS.md is . Your file appears as a dashed line.
What fills the file
Two things the agent can find by itself: a section that says what the project is, and a pasted directory tree. Bars show the share of files that have them, and the share of all tokens they take.
Files that kept the /init start
of CLAUDE.md files still open with the line /init writes. They have an overview section of the time, against for the rest. The line is found by text match, so a file that kept the line may have been edited since.
Two names, one file
Repos that have both a CLAUDE.md and an AGENTS.md. link one to the other, keep a near copy, and write two separate files.
Hooks: when they run
A hook is a command a repo asks the agent to run at a fixed moment, such as before a tool call. Only of repos commit one, so every share below is among repos that have hooks.
Commands people write down
Commands named in the files, grouped by toolchain. Each bar is the share of repos in that group that name the command.
What this cannot tell you
- It counts what is in the files, not whether the file helps an agent. A short file can be worse than a long one.
- Public repos that GitHub code search indexes, as of . Private repos are not here, and of the repos are small or personal, which is why popular repos are shown beside them.
- Your token count is an estimate: characters divided by , the median characters per token across the census files, measured with an o200k tokenizer. It is approximate, not Claude’s tokenizer, and it leaves out imports.
- Overview sections, directory trees and the
/initline use the same rules as the census. Commands, imports and “don’t” rules in the checker use simpler matching, so treat those three as approximate.
Method, in two sentences
The census sampled public repos through GitHub code search by file size, weighted each repo by how likely it was to be picked, and drew 95% intervals by resampling repos. Only counts and shares were kept, with no file text, repo names or paths.
Numbers are on the neri repository. The derived numbers are shared under CC BY 4.0.
Used in
- Loading…