Topic
Repository files such as CLAUDE.md and AGENTS.md that hold standing, project-specific instructions for AI coding agents.
Files seeded by Claude Code's /init scored a median 8.4 against 7.7 for other CLAUDE.md files in Reporails' scan of 8,220 files. Only about a fifth of an /init file is instruction, against close to three in ten elsewhere.
Reality
Claude Code 2.1.277 through 2.1.280 read AGENTS.md only when a remote feature flag resolved true, and 2.1.281 fixed it. Repos relying on the file got no project instructions with telemetry off or on Bedrock, Vertex, Foundry and gateways, according to a dev.to write-up.
Reality
Eval harness agents-md-evals found 25 of 26 assertions passed identically with or without a 755-line AGENTS.md file. The finding rests on that one file, and it assumes the agent loaded the file in the first place.
Reality
Rulestack found a skill file that named a deleted pnpm script for 24 days while 28 commit-time shape tests passed. A 26-line check against package.json exits 1 on any missing name, so a hook can block the commit that deletes a referenced script.
Reality
Claude Code started reading AGENTS.md on September 18, but only where no CLAUDE.md sits in the working directory or any parent above it. Monorepos need a stub file per package or a change to the instructionFiles setting.
Reality
Version 2.1.277 reads a shared AGENTS.md when no CLAUDE.md is in scope. The documented exceptions cover first sessions, telemetry-disabled runs, Bedrock, Vertex and Foundry, and the hooks governance tooling watches.
Reality
Anthropic shipped the AGENTS.md fallback in Claude Code 2.1.277 as a flag-gated built-in mod. Two machines on the same pinned version can load different project instructions. The same week's digest also logs a reverted deny-rule fix.
Perspective Coverage
4 publishersReality
Anthropic's memory documentation sets a 200-line target per instruction file and reserves real blocking for a hook. Everything written in markdown arrives as context the model weighs against your last message.
Reality
Claude can now read AGENTS.md, the file other agent tools already look for. XDA-Developers reports developers called the change small and long overdue, and the Claude-specific syntax that forced a second file has not moved.
Reality
Claude Code now supports AGENTS.md through a built-in plugin with four loading modes. The default treats the two filenames as alternatives, and the mode itself is set in user or managed settings outside the repo.
Reality
The rule telling a Claude Code session to gloss its own jargon sat unchecked in the repository for 68 days. Making it executable cost the team a narrower rule and one tightened parser.
Reality
A dev.to walkthrough splits one repo-wide Cursor rule into three scoped .mdc files for a Next.js App Router repo. The two path-scoped rules overlap on app/**, and a project that keeps routes under src/ matches neither.
Reality
A dev.to write-up cut a 194,492-character CLAUDE.md to 17,283 by changing what triggers each rule to load, sorting them by what breaks if one stays unloaded at the moment it matters. The harness that checks the result runs one headless Claude Code session per prompt.
Reality
A 558-file corpus puts gotchas last of nine categories at 13.6%, and the same study reports 32 to 38% recall on that slot, so the sturdier evidence for scarcity is a hand read of 120 pitfall entries.
Reality
The cap watched the file total, so when owner instructions pushed the rules body from 27,646 to 40,779 bytes in 18 days, the commit shortened its changelog entry to stay legal. The replacement test measures the body alone.
Reality
Three postures exist for an agent instruction file. Only the third holds at commit speed. The team running it watched its own CLAUDE.md reach 548KB before a context-measurement pass cut it to 34KB.
Reality