Two in three new open-source products now write instructions for coding agents
67% of the open-source products started since 2025 that we track keep an instruction file for coding agents (AGENTS.md, CLAUDE.md or similar), against 38% of those started in 2019 or earlier. In 16 days, 37% of the files were edited. Projects that keep one look like they pick different tools, but once project age and language are matched most of the gap disappears.
An instruction file — AGENTS.md, CLAUDE.md, Cursor or Copilot rules — is what a project writes for the coding agents that work in it: how to build, test and change the code. We check the repos of the open-source products we track for them. This month: how common they've become, what changed in 16 days, and whether the projects that keep them build differently. What goes inside the files is in our guide to them.
New projects keep one by default
817 of 1,545 actively developed product repos — 53% — keep at least one instruction file. Among projects started since 2025 it's 67% (370 of 556), and 71% of those started in 2026. Older projects added them too, but less often: 38% of those started in 2019 or earlier.
Of 1,545 product repos pushed to in the last 180 days, counted 2026-10-28.
They keep editing them: 37% changed in 16 days
We read the same 1,412 repos on October 12 and again on October 28. Of the 639 that had a main instruction file both times, 238 changed it. These aren't files written once at setup: over a third of them moved in two weeks.
The count barely moved — 725 to 734 repos with any file; 13 started keeping one and 4 stopped — but the direction is consistent:
| File | October 12 | October 28 | Added | Removed |
|---|---|---|---|---|
| AGENTS.md | 538 | 556 | 18 | 0 |
| CLAUDE.md | 424 | 423 | 8 | 9 |
| .claude folder | 315 | 321 | 9 | 3 |
| Cursor rules | 107 | 105 | 0 | 2 |
| Copilot instructions | 111 | 111 | 0 | 0 |
AGENTS.md grew and no repo dropped it; CLAUDE.md lost as many as it gained. Of the 13 repos that started keeping a file, 11 began with AGENTS.md. And among repos with both files, the share whose CLAUDE.md only points to AGENTS.md went from 53% to 55%. Two weeks is short; we'll know whether this is a trend at the next count.
Do these projects pick different tools?
Compared across all repos, the ones with instruction files look like a different crowd:
| Tool | Decision | With files | Without | Gap |
|---|---|---|---|---|
| Vitest | Testing | 61% | 38% | +23 |
| Better Auth | Authentication | 37% | 15% | +22 |
| Playwright | Testing | 49% | 33% | +17 |
| PostHog | Web Analytics | 57% | 41% | +16 |
| Hono | Backend Framework | 21% | 8% | +13 |
| Testing Library | Testing | 40% | 27% | +13 |
| Claude | LLM API | 52% | 40% | +11 |
| PostgreSQL | Database | 68% | 57% | +11 |
| React | Frontend Framework | 81% | 71% | +10 |
| SendGrid | Transactional Email | 18% | 31% | −13 |
But projects with instruction files are newer, and more of them are written in TypeScript, and both of those come with their own tools. Matched — only TypeScript projects started since 2024 — most of the gap goes away. Vitest, Playwright and React are used about as much by projects with and without the files.
A few gaps hold:
| Tool | Decision | With files | Without | Gap, matched | Gap, all repos |
|---|---|---|---|---|---|
| Better Auth | Authentication | 57% of 68 | 36% of 28 | +22 | +22 |
| PostHog | Web Analytics | 67% of 79 | 50% of 20 | +17 | +16 |
| Claude | LLM API | 65% of 86 | 49% of 35 | +17 | +11 |
| PostgreSQL | Database | 67% of 120 | 51% of 47 | +16 | +11 |
| SQLite | Database | 34% of 120 | 21% of 47 | +13 | +8 |
| Hono | Backend Framework | 47% of 96 | 35% of 34 | +12 | +13 |
| Testing Library | Testing | 45% of 193 | 34% of 62 | +11 | +13 |
| Vercel | Hosting | 52% of 143 | 63% of 43 | −10 | +9 |
| SendGrid | Transactional Email | 7% of 72 | 20% of 25 | −13 | −13 |
| Prisma | Database Access & ORMs | 20% of 85 | 38% of 29 | −18 | −4 |
TypeScript repos started since 2024; a repo counts for a decision when we can see what it picked. Gaps in percentage points.
Ahead in projects with instruction files: Better Auth, PostHog, Claude, PostgreSQL, SQLite, Hono and Testing Library. Behind: Vercel, SendGrid and Prisma. The comparison groups are small, though — 20 to 83 repos without instruction files per decision — so read these as signals to watch, not findings. One reverses: Supabase is +6 points across all repos but −5 matched; Vercel is +9 points across all repos but −10 matched.
To watch in November
- Whether AGENTS.md keeps growing while CLAUDE.md stays flat, and more CLAUDE.md files become pointers to it.
- Whether the matched gaps for Better Auth, PostHog, Claude hold as more repos are counted, or shrink like the others did.
- Whether a third of the files changing every two weeks is the norm or this fortnight's.
How this is counted
The sample is the GitHub repos of the products we track that someone pushed to in the last 180 days. For each, GitHub's API gives the root files, the instruction files coding agents read (AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules, .windsurfrules, .github/copilot-instructions.md) and the folders some agents read rules from (.cursor/rules, .claude, .github/instructions, .clinerules, .kiro/steering). The change compares the same 1,412 repos, active and not forks on October 12, read again on October 28; a file counts as edited when its text differs. A product's tools come from the evidence on its page, code included. Keeping an instruction file doesn't mean an agent wrote the product, and these are open-source projects only: closed-source products may differ. The files are their authors' writing, so we publish counts only.
Questions
How many open-source projects have an AGENTS.md or CLAUDE.md?
53% of the 1,545 actively developed product repos we track keep at least one instruction file for coding agents, and 67% of the ones started since 2025. AGENTS.md is in 617 of them, CLAUDE.md in 473.
Is AGENTS.md replacing CLAUDE.md?
In the same 1,412 repos between October 12 and October 28, AGENTS.md went from 538 to 556 with none removed, while CLAUDE.md stayed at 423. Of repos with both, 176 of 321 have a CLAUDE.md that only points to AGENTS.md.
Do projects built with coding agents use different tools?
Mostly not, once you compare like with like. Across all repos, those with instruction files use Vitest, Playwright, React far more, but among TypeScript projects started since 2024 the gap is under 5 points. A few gaps hold — Better Auth +22, PostHog +17, Claude +17 points — on small samples.