mcpField
A
agents-md-craft

agents-md-craft

github.com/robsonrungAgent workflows265 installsUpdated 2d ago
Repository

About agents-md-craft

agents-md-craft is a skill for creating, updating, auditing, and checking staleness in agent instruction files. It helps maintain concise root guidance grounded in verified project facts, with conditional detail kept in linked documents.

Key features

  1. 01Audit, create, update, and staleness checks
  2. 02Keep root instructions concise
  3. 03Move conditional guidance to linked files
  4. 04Verify commands and document pointers

Use cases

  1. 01Review an AGENTS.md or CLAUDE.md file
  2. 02Create project-specific agent instructions
  3. 03Restructure long guidance into focused documents

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/robsonrung/rar-skills --skill agents-md-craft

After installing, type/agents-md-craftin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

agents/
openai.yaml
references/
checklist.md
principles.md

Details

From the project README

Agent Instruction Files

Choose the branch before selecting a model. This workflow runs with the current coordinator and direct tools. It does not dispatch a worker.

Treat the root instruction file as cached context. Keep only rules that apply to most work. Move conditional detail to focused files and point to them from the root.

Outcome

Produce one authoritative instruction set grounded in verified project facts. An audit reports a proposed diff. A create or update request writes the agreed scope and verifies its pointers.

Workflow

  1. Classify the request as audit, create, update, or staleness check. Inspect the target files and the project facts that support their instructions: purpose, layout, commands, and durable constraints.
  2. If both AGENTS.md and CLAUDE.md exist, choose the maintained file as canonical unless the project requires separate content. Keep the other as a short text pointer when the active tools can follow it. Do not create two copies that will drift.
  3. Keep the root file to its smallest coherent shape:
    1. project purpose and major areas;
    2. commands and constraints that apply to normal work;
    3. short links to conditional guidance.
  4. Move runbooks, long command recipes, subsystem rules, and ADRs into agent_docs/ or the project's existing documentation. Point to existing formatters, linters, and hooks for rules they enforce. Preserve team standards and security boundaries that need prose. Use stable file or symbol pointers instead of copied code; verify line numbers when including them.
  5. For audit, return findings, a budget view, and a concrete draft without writing files. For create or update, write only the requested files and their necessary agent_docs/ references. Do not invent project facts.
  6. Check changed commands against their actual definitions and resolve changed links or line pointers. Run commands only when needed to resolve a material uncertainty or satisfy a repository gate; documenting a deployment command does not authorize deployment. Reuse valid evidence for unchanged content. Report the canonical file, changes, unresolved facts, and validation evidence.

Read references/principles.md when deciding placement or scope. Use references/checklist.md for an audit or a substantial restructure; a small edit needs only the checks for its changed contract.

Comments

Markdown supported