agents-md

name

agents-md

description

Make AGENTS.md the canonical project-instructions file and ensure CLAUDE.md and GEMINI.md are thin pointers that @-import it. Promotes existing CLAUDE.md/GEMINI.md content into AGENTS.md when AGENTS.md is missing or weaker, creates stub CLAUDE.md and GEMINI.md when absent, and bootstraps a fresh AGENTS.md when the project has no instructions file at all. Once canonical, also audits an existing AGENTS.md for structure drift, verifies its build/test/lint commands and gotchas are still accurate, and harvests durable lessons from the current session. Use when the user wants to tidy up agent instructions, onboard a project, check whether AGENTS.md is up to date, audit its structure, add what was learned this session, or says /agents-md.

Agents-MD

Reconcile project-instructions files so AGENTS.md is canonical and CLAUDE.md / GEMINI.md are thin stubs that @-import it. The same project then works from Claude Code, Gemini CLI, GitHub Copilot, or any agent that honours AGENTS.md, with no duplicated content.

Two modes

Determine the mode first, using the same inventory the Decision table needs (see "Inputs").

• Reconcile/Bootstrap (default) — AGENTS.md is missing, weaker than an existing CLAUDE.md/GEMINI.md, or the stubs don't import it. Run the Decision table. One-shot.
• Update/Audit — AGENTS.md exists, is canonical, and both stubs already import it (the content | stub | stub state). Run "Update mode". Re-runnable.

Flags

• --learn — skip straight to the session-lessons harvest. Bypass mode determination, the Decision table, and the structure/freshness jobs; run only job 3 of Update mode against the existing AGENTS.md. Use at a session's end when you only want to record durable lessons, not a full audit. If AGENTS.md does not exist, report that and fall back to normal mode determination.

Why this works

• Claude Code reads CLAUDE.md, not AGENTS.md, but CLAUDE.md supports @path imports, and Anthropic's docs recommend @AGENTS.md as the canonical pattern. (See https://code.claude.com/docs/en/memory#agents-md.)
• Gemini CLI reads GEMINI.md and supports @file.md imports with relative or absolute paths. (See https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/gemini-md.md.)
• GitHub Copilot reads AGENTS.md natively — the nearest one in the directory tree wins — so it needs no stub. Its legacy .github/copilot-instructions.md accepts only inline Markdown with no import syntax, so it can't point at AGENTS.md; treat it as a content source to drain, not a stub. (See https://docs.github.com/en/copilot/customizing-copilot/adding-repository-custom-instructions-for-github-copilot.)

So CLAUDE.md → @AGENTS.md and GEMINI.md → @AGENTS.md give both tools the same content without duplication, while Copilot picks up AGENTS.md directly.

Bootstrap template (the rubric)

This is both the template for fresh files and the scoring rubric for choosing between existing files. A good AGENTS.md covers exactly three things:

1. WHY — what is this project and what problem does it solve?
2. WHAT — repo structure, language boundaries, key entry points.
3. HOW — commands to build/test/lint, plus gotchas not discoverable from code alone.

Point to docs; don't repeat them. Everything else (architecture, API surfaces, coding style) is discoverable via tools, MCPs, skills, and reading the code. Keep it under 200 lines — both tools load these files in full at session start, and longer files reduce adherence.

Inputs

Operate on the project root where the skill is invoked. The relevant files:

• AGENTS.md (canonical target)
• CLAUDE.md (Claude Code pointer)
• GEMINI.md (Gemini CLI pointer)
• .github/copilot-instructions.md (GitHub Copilot legacy source — a content source to consolidate, never a stub; see "GitHub Copilot")

"Substantive content" means more than boilerplate or an import line. A file whose only non-whitespace content is @AGENTS.md is a stub, not content.

Decision table

Inventory what is present and substantive, then follow the matching row.

AGENTS.md	CLAUDE.md	GEMINI.md	Action
none	none	none	Bootstrap — create AGENTS.md from scratch using the rubric, then stub both pointers
content	none	none	Create stub CLAUDE.md and GEMINI.md that import AGENTS.md
content	stub	stub	Update/Audit — already canonical; run "Update mode" below
content	content	any	Score and merge — see "Picking the canonical source" below
content	any	content	Score and merge — see "Picking the canonical source" below
none	content	none	Score CLAUDE.md against the rubric; promote (or rewrite to match) → AGENTS.md, then stub both pointers
none	none	content	Score GEMINI.md against the rubric; promote (or rewrite to match) → AGENTS.md, then stub both pointers
none	content	content	Score both against the rubric; pick the one that scores higher as the base, backfill missing sections from the other, then stub both

.github/copilot-instructions.md sits outside this table — it is never a stub. If it holds substantive content, fold it into AGENTS.md per "GitHub Copilot" below; otherwise ignore it.

Picking the canonical source

When more than one file is in the running — including a substantive .github/copilot-instructions.md — score each candidate against the rubric:

1. Rubric coverage — WHY, WHAT, HOW. Each present, accurate, and non-stale scores a point; missing or wrong scores zero.
2. Orientation focus — points to docs rather than duplicating them, and stays within ~200 lines.
3. Freedom from tool-specific noise — generic guidance (structure, build commands, conventions) scores higher than tool-specific features (Claude skills, plan mode, Gemini checkpointing/sandbox, Copilot coding-agent settings). Tool-specific content is fine, but lives in the stub, not in AGENTS.md.

The highest-scoring file becomes the base. Backfill missing rubric sections from the others. Peel anything tool-specific into the matching stub under ## Claude Code or ## Gemini CLI; Copilot-specific guidance stays in .github/copilot-instructions.md (see "GitHub Copilot"). If two candidates score roughly equal, show the user a short diff summary and ask which should be canonical before writing.

Stub templates

CLAUDE.md

The import line is the only required content. Append any Claude-specific guidance after the import under a dedicated heading.

@AGENTS.md

With Claude-specific additions:

@AGENTS.md

## Claude Code

<Claude-specific notes here>

GEMINI.md

Gemini's import example uses the @path form; use @AGENTS.md for parity with CLAUDE.md.

@AGENTS.md

With Gemini-specific additions:

@AGENTS.md

## Gemini CLI

<Gemini-specific notes here>

GitHub Copilot

Copilot has no stub. It reads AGENTS.md natively, so once content is canonical Copilot picks it up with no extra file. Handle .github/copilot-instructions.md by what it holds:

• Substantive content — score and merge it into AGENTS.md like any other candidate (see "Picking the canonical source"), then reduce the file to genuinely Copilot-specific guidance or delete it. It cannot @-import AGENTS.md, so never leave an import line in it.
• Absent or already thin — leave it; native AGENTS.md support covers Copilot.

If an org mandates the file exist, make it a short prose pointer plus any Copilot-specific notes — a pointer in prose, not an import:

The canonical project instructions live in `AGENTS.md`; read that file.

<Copilot-specific notes here>

Update mode

The file is already canonical. Run all three jobs in sequence, then gate all proposed changes behind a single confirmation — write nothing before the user approves. With --learn, run only job 3 and skip jobs 1 and 2 (still gate behind one confirmation).

1. Structure audit — score the existing AGENTS.md against the rubric (above); do not restate it. Flag missing or malformed WHY/WHAT/HOW sections, length creeping past ~200 lines, and tool-specific content that has leaked in from a stub (it belongs under ## Claude Code / ## Gemini CLI). Also flag substantive content that has accumulated in .github/copilot-instructions.md; it should be folded back into AGENTS.md (see "GitHub Copilot"). Propose concrete fixes.
2. Freshness check — verify the WHAT/HOW claims against the repo using read-only tools. Confirm the build/test/lint commands exist (cross-check Makefile, package.json scripts, or equivalent), that documented entry points and structure still match, and that listed gotchas still apply. Flag anything stale or wrong, and cite what you checked.
3. Session-lessons harvest — review the current session for durable, non-obvious facts that fit WHY/WHAT/HOW: a gotcha that cost time, a non-obvious command, a constraint not discoverable from the code. Exclude anything a reader could learn from the code. If nothing qualifies, skip — do not force it.

Then present the combined diff or a short summary and write only after the user confirms. Keep AGENTS.md under ~200 lines and tool-agnostic. After writing, run the Verification checklist.

Verification

After writing, confirm:

1. AGENTS.md matches the rubric (WHY / WHAT / HOW, under ~200 lines, tool-agnostic).
2. CLAUDE.md begins with @AGENTS.md on its first non-empty line.
3. GEMINI.md begins with @AGENTS.md on its first non-empty line.
4. .github/copilot-instructions.md, if present, holds no substantive content missing from AGENTS.md and contains no import line.
5. No content lives only in CLAUDE.md, GEMINI.md, or .github/copilot-instructions.md unless it is genuinely tool-specific.

Report the final state to the user as a short summary — which file is canonical, what was moved, what was created.

Guidelines

• Reconcile/Bootstrap is one-shot — don't re-run it once the three files are canonical. Update mode is the re-runnable path: use it when you suspect AGENTS.md has drifted from the repo, or at the end of a session that surfaced durable lessons.
• Preserve the user's voice. When merging, keep original wording rather than rewriting for style.
• Never delete content without confirming the user is okay losing it. If in doubt, keep both versions in AGENTS.md and flag the overlap.
• Wrap all Markdown output at 80 columns.