Modern AI coding tools each ship with their own per-project configuration file. The Opus 4.7 corpus gives us a rare quantitative comparison.
The configs
| File | Repos | Tool |
|---|---|---|
| CLAUDE.md | 2,804 | Claude Code |
| AGENTS.md | 999 | Claude Code / multi-agent |
| .cursorrules | 56 | Cursor |
| .windsurfrules | 25 | Windsurf |
| .codex | 16 | OpenAI Codex |
| CODEX.md | 6 | OpenAI Codex |
| AI.md | 3 | Generic |
| claude-rules.md | 1 | Claude (non-canonical) |
The ratios
- CLAUDE.md appears in 2,804 repos
- .cursorrules appears in 56 repos
- Ratio: 50.1×
Even combining every non-Claude AI config (.cursorrules + .windsurfrules + .codex + CODEX.md + AI.md + claude-rules.md = 107 repos), CLAUDE.md alone outnumbers them 26×.
Keep in mind: this is the Claude Opus 4.7 corpus specifically, which should bias toward Claude files. But the ratio would still be striking in a neutral corpus — Cursor, Windsurf, and Codex just haven’t established a comparable file-level footprint on GitHub.
Why CLAUDE.md dominates
Three reasons:
1. Claude Code made it a convention early
Claude’s docs have called out CLAUDE.md since mid-2024 as the canonical way to give agents project context. Cursor relied on their IDE settings for longer before standardizing on .cursorrules.
2. It’s markdown, not dotfile
CLAUDE.md is human-readable, version-controlled, visible in GitHub file browsers, and ranks in GitHub search. Dotfiles (.cursorrules) are hidden by default and harder to notice/copy from other projects.
3. AGENTS.md extends the pattern
CLAUDE.md’s natural sibling AGENTS.md (999 repos) creates a two-file pattern for multi-agent workflows. This makes the “Claude memory file” system more extensible than a single .cursorrules.
What it means for the ecosystem
Three implications:
For Anthropic (Claude)
CLAUDE.md has become a de-facto open standard. Other AI tools that want agent persistence would benefit from adopting this convention rather than creating their own silo (which .cursorrules and .windsurfrules attempted).
For tooling vendors
If you’re building a tool that plugs into the AI coding ecosystem (CI, linting, static analysis, etc.), check for CLAUDE.md first. It’s the highest-probability signal that a repo is AI-assisted and has agent context.
For developers
Writing a good CLAUDE.md is becoming a skill. The CLAUDE.md Phenomenon covers what’s typically inside one. If you’re starting a new project, write one on day one — it’s the single highest-ROI investment for AI collaboration.
The Codex footprint
Codex’s markers are tiny: 16 repos with .codex, 6 with CODEX.md. These are specifically OpenAI Codex CLI markers.
This tracks with Codex’s adoption curve — still early, users are mostly on the CLI which doesn’t (yet) require a dedicated config file. Contrast with Claude Code, which has been prompting users to write CLAUDE.md for months.
The Windsurfrules gap
Windsurf (by Codeium) has 25 repos with .windsurfrules. That’s a fifth of Cursor’s count. Reasonable given Windsurf is newer, but also a signal — even among users of newer AI IDE tools, persistent project-config files aren’t yet standard practice.
Recommendation
When building any AI-coder tooling in 2026, treat CLAUDE.md as the canonical project-context file. It has:
- Highest adoption (50× ratio vs alternatives)
- Best documentation culture around it (READMEs + guides)
- Strong positive correlation with higher-quality repos
- Markdown format = universally parseable
The fragmentation across .cursorrules / .windsurfrules / .codex is real, but the gravitational pull toward CLAUDE.md suggests it will become the ecosystem default.
Related: The CLAUDE.md Phenomenon.