| name | agents-md-editor |
| description | Use when creating, editing, or reviewing AGENTS.md, agents.md, CLAUDE.md, Copilot instructions, Cursor rules, or any always-read repository instruction file. Keep persistent agent instructions small, durable, non-duplicative, and action-oriented; classify content that belongs in commands, onboarding, reference docs, handoff state, enforcement, or removal instead of preserving it in always-read files. |
Agents MD Editor
Purpose
Keep always-read instruction files small, durable, and action-oriented.
Treat these files as startup instructions for future agent sessions, not as session notes, decision logs, onboarding documents, or general reference documents.
Core Rule
Keep only content that should change agent behavior in future sessions.
Good content usually affects one of these:
- command choice
- safety boundaries
- repository-specific workflow
- where to look before acting
- project-specific conventions not obvious from code or tooling
Classification
For each candidate line, choose one destination:
- Keep: durable behavior rule, safety boundary, workflow rule, or short routing instruction.
- Commands: recurring install, build, test, generation, or validation command agents need during normal work.
- Onboarding: first-run setup, local bootstrap, or human-only environment preparation.
- Reference: architecture, naming, ownership, repo/service boundaries, or durable facts better explained elsewhere.
- Handoff: current state, temporary migration status, next step, open work, blockers, or session-specific context.
- Enforcement: scripts, hooks, CI, tests, linters, type checks, generated checks, or policy already enforced mechanically.
- Remove: duplicate, stale, vague, historical, rejected alternative, decision trace, or meta-commentary.
Use the repository's existing surfaces for non-AGENTS content. Do not invent a new documentation structure unless the user asks.
Editing Rules
- Prefer positive instructions.
- Use negative instructions only for durable safety, security, ownership, or destructive-change boundaries.
- Do not preserve temporary state just because it mattered during the edit.
- Do not record rejected alternatives, reversals, or "we decided not to" history in always-read instruction files.
- Do not restate what tooling already enforces unless agent behavior before tooling runs would otherwise be unsafe.
- Replace detailed reference facts with a short pointer when agents only need to know where to look.
- Keep recurring commands when they are needed for normal agent work.
- Move first-run setup and local environment bootstrap instructions to onboarding or setup surfaces.
- Keep the file short enough to scan before acting.
- Avoid vague rules such as "write clean code", "follow best practices", or "be careful".
Duplicate Check
Before saving, compare proposed content against:
- existing always-read instruction file content
- README
- docs
- scripts
- CI configuration
- existing skills or rule files
If the same instruction already exists elsewhere, keep only a short pointer when that pointer changes agent behavior.
Review Before Saving
Before editing an always-read instruction file, produce a short review with:
- lines to keep
- lines to move and their destination
- lines to remove
- duplicate or stale content found
- proposed diff
Do not save changes until the user approves unless the user explicitly requested direct editing.
Final Gate
Before saving, every remaining line must answer yes to at least one question:
- Should this change how the agent acts in future sessions?
- Does this prevent a likely unsafe or destructive action?
- Does this tell the agent which command to run for normal work?
- Does this tell the agent where to look before acting?
- Is this repository-specific and not obvious from code, tooling, or existing docs?