agents-md-improver（MD 改良剂）

AGENTS.md improver

This skill helps keep repository instructions accurate and consistent. When a user corrects the coding agent's behavior, treat it as a potential update to the repo's instruction files and propose a clean, integrated edit to AGENTS.md.

Usage

Use this skill anytime the user:

• Corrects the agent ("don't do X", "do Y instead", "from now on...", "remember to...", "add a rule that...").
• Asks to update instruction files like AGENTS.md, CLAUDE.md, .claude/CLAUDE.md, or GEMINI.md, or to align or consolidate those files.
• Notices contradictions, duplication, or drift between repository instruction files.

Requirements

• Repo-local file access to read and edit AGENTS.md.
• If present in the repo, read access to CLAUDE.md and/or .claude/CLAUDE.md.

What It Does

• Proposes a patch to incorporate the user's correction into AGENTS.md seamlessly.
• Looks for contradictions and resolves them in the proposal (or asks for clarification when needed).
• If repo-local CLAUDE.md or .claude/CLAUDE.md exists, proposes consolidating instructions into AGENTS.md, then offers follow-up options to delete, symlink, or stub those files.

How It Works

Scope

This skill is repo-specific.

• Only consider repo-local instruction files:

- AGENTS.md - CLAUDE.md (repo root) - .claude/CLAUDE.md - GEMINI.md (repo root and/or repo subdirectories if present)

• Ignore system-wide instructions and files under the user's home directory (for example ~/.claude/, ~/.gemini/, ~/.config/, etc.).

Propose-first workflow (do not auto-edit)

1. Read AGENTS.md.
2. Identify the user's correction and translate it into a durable, repo-specific instruction.
3. Decide where it belongs:

- Prefer editing the most relevant existing section. - If it introduces a genuinely new topic, propose a new ## section near related content.

1. Draft the exact text change(s) as a small patch/snippet.
2. Ask for confirmation before applying:

- "Apply these changes to AGENTS.md? (y/n)"

Only apply edits after the user confirms.

Seamless integration rules

• Do not add dated changelog entries like "YYYY-MM-DD: correction".
• Update the existing prose/bullets/steps so the rule reads like it has always been part of the document.
• Keep formatting consistent with the surrounding AGENTS.md style.

Contradiction checks

Before proposing a patch:

1. Scan AGENTS.md for conflicts with the new correction.
2. If a conflict exists, include conflict resolution in the proposed patch by rewriting/removing the conflicting lines.
3. If resolution is ambiguous, ask a single targeted question before proposing a patch.

Then scan repo-local guidance (if present):

• Compare AGENTS.md against CLAUDE.md and .claude/CLAUDE.md.
• Flag duplication and conflicts.
• Prefer consolidating into AGENTS.md as the canonical source of truth.

Consolidation flow for CLAUDE.md files (if present)

If CLAUDE.md and/or .claude/CLAUDE.md exists in the repo:

1. Propose a consolidation patch to AGENTS.md.
2. After the user approves the AGENTS.md patch, propose what to do for each CLAUDE file:

- Delete the file. - Replace the file with a symlink pointing to AGENTS.md. - Replace the file contents with a small stub that points humans/tools to AGENTS.md (fallback when symlinks are undesirable).

1. Confirm destructive actions (deletes) in a separate prompt.

Symlink directions (when chosen):

• CLAUDE.md -> AGENTS.md
• .claude/CLAUDE.md -> ../AGENTS.md

Notes for multi-instruction environments

Some harnesses merge multiple instruction sources and may apply conflicting rules non-deterministically. Prefer a single canonical source (AGENTS.md) and reduce parallel instruction files via consolidation + delete/symlink/stub.