AGENTS.md vs CLAUDE.md vs Cursor Rules: One Source of Truth

Teams now use several coding agents, and duplicated instruction files quickly disagree. AGENTS.md can hold shared repository guidance while thin tool-specific files keep only behavior unique to each agent.

When to use

Use the same commands and constraints across multiple coding agents.
Prevent copied instruction files from drifting.
Keep scoped or tool-only features in their native configuration.

Example

Codex, OpenCode and Cursor share commands and safety rules from AGENTS.md; CLAUDE.md imports that source and adds one Claude-only review skill.

TL;DR

Write rules once in AGENTS.md; make CLAUDE.md import it.

bash

echo '@AGENTS.md' > CLAUDE.md

Steps

1
Put shared rules in AGENTS.md
Start with commands and durable repository constraints that should remain identical regardless of which supported agent performs the work.
AGENTS.md
```
# Agent instructions

## Commands
- Install: `npm ci`
- Test: `npm test`

## Rules
- Never edit src/gen/.
- Add a test for every bug fix.
```
2
Point Claude Code at it
Keep CLAUDE.md thin: import the shared source, then add only behavior that applies specifically to Claude Code.
CLAUDE.md
```
@AGENTS.md

## Claude-only
- Use the review-diff skill before finishing.
```
3
Keep Cursor rules for scoping only
Use .mdc files for glob-specific rules, not for repeating AGENTS.md.
4
Personal defaults
Keep preferences that should not affect teammates in each tool's personal instructions directory rather than committing them to the repository.
paths
```
~/.codex/AGENTS.md     # Codex, all projects
~/.claude/CLAUDE.md    # Claude Code, all projects
```

Gotchas

Two copies of the same rule will drift; delete duplicates.
Nested AGENTS.md files override parent ones for that folder.
Tool-specific features (skills, subagents, Bugbot) stay in their own files.

Cheat sheet

AGENTS.md	Codex, OpenCode, Cursor
CLAUDE.md + @AGENTS.md	Claude Code
.cursor/rules/*.mdc	Cursor scoped rules
.cursor/BUGBOT.md	Cursor Bugbot reviews

Sources

Reviewing agent output? CodeCrab reviews pull requests locally with your own AI tools.