---

name: agents-md-improver description: "Maintain AGENTS.md-based project instructions for tools that support AGENTS.md. Use for instruction audits, session learnings, AGENTS.md updates, or migrating useful CLAUDE.md rules into a shared instruction source. Not for general docs, code review, runtime debugging, or global memory unless asked."

AGENTS.md Improver

Overview

Maintain concise project instructions for AGENTS.md-compatible tools such as Codex, Copilot, Cursor, and similar coding agents.

Core principle: Find the effective instruction source -> extract durable learnings -> propose compact diffs -> apply approved changes only.

Announce at start: "I'm using the agents-md-improver skill to maintain concise AGENTS.md project instructions."

Prerequisites

• A repository or workspace with AGENTS.md, nested AGENTS.md, or compatible instruction files.
• A user request to audit, improve, revise, migrate, or capture learnings into project instructions.

Workflow

Phase 1: Identify Instruction Targets

Find instruction files from the repository root:

find . \( -name "AGENTS.md" -o -name "CLAUDE.md" -o -name ".claude.local.md" \) 2>/dev/null | head -50

Use these targeting rules:

• ./AGENTS.md: primary shared project instructions for AGENTS.md-compatible tools.
• Nested AGENTS.md: module-specific instructions; update only when the learning applies to that subtree.
• Include-only AGENTS.md: if it contains only an include such as @./CLAUDE.md, read the target and treat that target as the effective source. Update the target only when the repository intentionally shares one instruction file across tools.
• CLAUDE.md: Claude Code-specific instructions, or a shared source when AGENTS.md intentionally points to it. Do not update it just because it exists.
• .claude.local.md: personal Claude Code local notes. Do not create or edit it for shared AGENTS.md project memory.
• User/global instruction files, such as ~/.codex/AGENTS.md: update only when the user explicitly asks for global behavior.
• .project_context/: historical exploration, plans, and execution summaries. Use it for trace records, not agent steering instructions.

If no AGENTS.md exists, propose creating the narrowest useful one, usually ./AGENTS.md.

Phase 2: Choose The Mode

Choose exactly one primary mode before scanning or proposing edits:

• Repository Audit Mode: use when the user asks whether instructions are stale, incomplete, inaccurate, outdated, or should be refreshed against the current repository. Then read references/repository-audit.md.
• Session Learning Mode: use when the user asks to capture this session's learnings, revise instructions from the current conversation, or preserve newly discovered repo-specific behavior. Then read references/session-learning.md.
• Migration Mode: use when moving useful guidance from CLAUDE.md or another tool-specific instruction file into an AGENTS.md-compatible source. Reuse Session Learning Mode for extraction rules, but keep tool-only slash commands, hooks, or UI shortcuts in tool-specific files unless the user asks to make them shared.

Phase 3: Draft Compact Updates

AGENTS.md is prompt context. Prefer dense, operational bullets.

Good format:

- Use `rg` for repo-wide search; fall back only if unavailable.
- Run `pnpm test -- --runInBand` for flaky integration tests in this repo.
- Keep API compatibility normalization in the service layer unless a spec explicitly changes request validation.

Avoid adding:

• generic engineering advice;
• one-off debugging details unlikely to recur;
• unverified assumptions or stale guesses;
• long explanations, logs, secrets, tokens, private URLs, or machine-specific credentials;
• duplicates of existing instructions.

Phase 4: Propose Before Editing

Always show proposed changes before modifying files. Group proposals by target file.

Use this format:

### Update: ./AGENTS.md

Why: <one-line reason this helps future agent sessions>

```diff
+ <concise addition>
```

If multiple files could be updated, explain the placement tradeoff:

• root AGENTS.md for project-wide behavior;
• nested AGENTS.md for module-local behavior;
• .project_context/ for historical records;
• CLAUDE.md only for Claude Code-specific behavior or an intentional shared source.

Ask for approval before editing. If the user already explicitly said to apply changes, proceed with the smallest safe patch.

Phase 5: Apply Carefully

When approved:

• use minimal patches and preserve the file's existing organization;
• insert under an existing relevant heading when possible;
• create a short new heading only if no suitable section exists;
• do not reorder unrelated content;
• do not remove or rewrite existing instructions unless the user asked for cleanup or the instruction is clearly obsolete;
• after editing, re-read the changed section and report exactly what changed.

Common Mistakes

• Treating .project_context/ as agent steering context. It is for traceability, not instructions that should load every session.
• Updating CLAUDE.md for a shared AGENTS.md request when the repo does not intentionally share one source file.
• Missing an include-only AGENTS.md and patching the wrapper instead of the effective source.
• Capturing logs, secrets, private URLs, or one-off debugging notes as durable instructions.
• Adding broad best practices that any competent agent already knows.

Output

For Repository Audit Mode, provide the quality report format from references/repository-audit.md.

For Session Learning Mode or Migration Mode, provide:

• files changed;
• concise summary of additions;
• any skipped proposals and why.

References

• repository-audit.md - Repository scan scope, quality criteria, and report template for stale or incomplete instructions.
• session-learning.md - Session learning extraction rules and concise update proposal template.