AI Code Pulse
AI adoption#
Every active repository scanned had at least one AI configuration file. The average ships 2.6.
| Config file | Share of repos |
|---|---|
| CLAUDE.md | 54% |
| AGENTS.md | 49% |
.claude |
42% |
.cursor |
40% |
.cursorrules |
40% |
| copilot-instructions.md | 22% |
.aiderignore |
2% |
| CODEX.md | 1% |
The most instrumented repositories — PostHog, Superset, Ghost, cal.com — ship five or more AI config files each.
Vendor lock-in is not happening#
46% of repositories configure two or more AI vendors, averaging 2.2 configs per repository. Claude leads at 81% adoption, Cursor follows at 40%, GitHub Copilot at 22%, CodeRabbit at 9%. The default is not picking one tool. It is shipping many.
What teams tell their AI to care about#
| Topic | Share of instruction files |
|---|---|
| Testing | 94% |
| Code quality | 91% |
| Architecture | 85% |
| API design | 80% |
| Database | 77% |
| Documentation | 75% |
| Security | 69% |
| Performance | 49% |
The median instruction file is 3,500 bytes, roughly 100 lines of guidance. 22% are stubs: the config exists but says nothing. Teams invest most in telling AI how to test and least in telling it how to optimize.
The documentation paradox#
Framework repositories ship 25 times more documentation than business repositories — 4,040 KB and 341 docs on average, against 150 KB and 15 docs. Review accuracy runs the other way: 81% for business repositories, 68.5% for frameworks.
Business repositories explain why. Frameworks document what. The reviewer learns more from intent than from API surface.
The YC effect#
Across 43 YC-backed repositories and 240 non-YC repositories in the same ecosystem, YC companies adopt AI tooling at 1.6 times the rate: 56% ship some AI config against 34%, 30% ship CLAUDE.md against 18%, and 26% ship a Cursor config against 11%.
Copilot inverts the pattern — 10% among non-YC repositories against 2% for YC. YC companies pick best-of-breed tools; everyone else uses what ships with the IDE.
Related#
- Product drift — what the review engine looks for
- Writing specs Prelint can review — why intent beats API surface
- Documentation home