Jev read every CLAUDE.md
What 25,784 CLAUDE.md files actually say to the agent.
Claude Code started reading AGENTS.md last week. GitHub now has over half a million public CLAUDE.md files. TypeSafe's Jev, a decision model that returns probabilities instead of prose, read a size-stratified sample of them and answered eleven questions about each. Then it can read yours.
25,784
CLAUDE.md files read
sampled from 547,992 public ones
35%
start with the /init boilerplate
17% look never edited
32%
tell the agent to verify its work
57% give exact commands
48%
contain a hard “never”
1% shout in caps
9%
read as LLM-written
0.9% contain a real token or DB password
What files contain
Share of files where Jev put the probability above 50%.
- Exact build / test / lint commands57% · 47%
- Tells the agent to verify its work32% · 48%
- Maps where things live in the codebase69% · 52%
- Explicit prohibitions (never, do not)48% · 67%
- Git rules: commits, branches, pushing6% · 9%
- Bans comments or explanatory prose0% · 1%
- Reads as LLM-generated9% · 6%
- Shouts in ALL CAPS1% · 1%
- Leaks something that should be private4% · 2%
Second number: 10,662 AGENTS.md files, for comparison.
What kind of document it is
- Architecture map40%
- Ops manual39%
- Style guide9%
- Persona7%
- Stub3%
- Todo dump1%
Does length help?
| size | files | commands | verify | /init |
|---|---|---|---|---|
| <1k | 2,178 | 19% | 17% | 6% |
| 1-2.5k | 3,781 | 55% | 32% | 26% |
| 2.5-5k | 5,957 | 72% | 39% | 46% |
| 5-10k | 4,474 | 65% | 34% | 43% |
| >10k | 9,394 | 47% | 27% | 31% |
Grade yours against the 25,784
one Jev call, about a tenth of a cent, on your own keyFree at console.typesafe.ai/keys, $5 credit included. The key is forwarded per request and never stored or logged.