Agents md init

AGENTS.md Init

Use this skill to analyze a project and create practical, current instructions for coding agents. Treat AGENTS.md and CLAUDE.md as equivalent instruction files: prefer AGENTS.md as the canonical shared file and make CLAUDE.md import it with @AGENTS.md unless the user explicitly wants another setup.

When To Use

• The user asks for /init-style project instructions, project memory, or a repository guide for AI coding agents.
• The project has no useful AGENTS.md or CLAUDE.md, or the user wants to replace a starter/template file with project-specific guidance.
• The user asks to support both Codex-style AGENTS.md and Claude-style CLAUDE.md instructions.

If the user asks to edit an existing mature file, use the maintenance workflow instead.

Load References

Read references/instruction-file-practices.md when you need the quality rubric, recommended sections, import/symlink rules, or source links.

Related Skills

When phrasing rules in the file, also use the prompt-craft skill for general prompting principles: motivating constraints, "when X, do Y" rules, output contracts, and anti-patterns. AGENTS.md and CLAUDE.md are additions to the system prompt, so prompting guidance applies.

Initialization Workflow

1. Discover current state

   • Check for AGENTS.md, CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, AGENTS.override.md, .claude/rules/, nested instruction files, and relevant .gitignore entries.
   • Determine whether CLAUDE.md imports @AGENTS.md, is a symlink, is duplicated content, or is intentionally separate.
   • Do not overwrite a non-empty existing instruction file without preserving its content or getting explicit approval.

2. Analyze the project

   • Identify languages, frameworks, package managers, task runners, test frameworks, build tools, lint/format tools, CI workflows, deployment hints, and monorepo/package boundaries.
   • Prefer authoritative files: README*, package manifests, lockfiles, Makefiles, task configs, CI files, Docker files, test configs, and existing docs.
   • Run cheap read-only inspection commands first. Run validation commands only when they are safe and appropriate for the repo.

3. Choose the file plan

   • Default: create AGENTS.md at the repository root and create CLAUDE.md containing @AGENTS.md, followed by any Claude-specific notes.
   • If CLAUDE.md exists but AGENTS.md does not, create AGENTS.md from the existing content plus project analysis, then reduce CLAUDE.md to an import wrapper with any truly Claude-specific content preserved.
   • If both files exist but are not aligned, inspect both, merge universal guidance into AGENTS.md, and keep CLAUDE.md as an import wrapper unless the user wants separate files.
   • Use a symlink only when the user explicitly prefers it, there is no Claude-specific content to add, and the operating system supports symlinks without elevated privileges. On Windows, use @AGENTS.md instead.
   • Create separate full files only when the user asks or a tool requires materially different syntax. Keep shared guidance in AGENTS.md and isolate file-specific notes in small labeled sections.

4. Draft AGENTS.md

   • Start with the project name and a one-paragraph purpose only if it helps an agent orient quickly.
   • Include only relevant sections and target under 200 lines for each loaded instruction file. Favor a compact core: Project Overview, Repository Layout, Setup, Development Commands, Testing, and Architecture Notes. Add Security, Environment, Workflow, or Known Gotchas only when they contain project-specific guidance.
   • Use exact commands from the repo, with short explanations of when to run them. Mark commands as required, preferred, expensive, or optional when that distinction matters.
   • Document durable project-specific behavior, not one-off task notes or generic advice every coding agent should already know.
   • Put universal instructions first. Add AGENTS.md-specific or CLAUDE.md-specific notes only when the tools truly differ.
   • For Claude-only instructions that apply only to certain paths, prefer .claude/rules/ with paths frontmatter instead of bloating the main file.

5. Create the files

   • Write AGENTS.md.

   • Unless the user requested separate files, write CLAUDE.md as an import wrapper:

     @AGENTS.md
     
     ## Claude Code
     
     [Claude-specific notes, only when needed.]
     

   • If running inside Claude Code, this skill replaces or extends the built-in /init flow with a portable AGENTS.md-first variant. Claude Code's built-in /init can also read an existing AGENTS.md, and CLAUDE_CODE_NEW_INIT=1 enables its interactive multi-phase init flow.

6. Verify

   • Confirm both paths resolve and contain the intended content:

     test -s AGENTS.md
     test -s CLAUDE.md
     grep -q '^@AGENTS.md$' CLAUDE.md || test -L CLAUDE.md
     

   • Validate every documented command as far as feasible. If a command was not run, say why.

   • Run git diff -- AGENTS.md CLAUDE.md or equivalent and review the result before finalizing.

Writing Rules

• Be concise, specific, and operational. A short accurate file beats a broad policy document. The file loads into the system prompt every session, so each line must solve an observed problem, not a theoretical one.
• Use Markdown headings and bullets. Keep commands in backticks or fenced code blocks.
• Prefer "when X, run Y" instructions over personality traits or slogans.
• State each rule once. Current frontier models follow instruction contracts closely; contradictory or redundant rules destabilize behavior more than missing detail, and a brief instruction steers a whole behavior class.
• Reserve absolute language (ALWAYS, NEVER, must) for true invariants such as safety rules and prohibited actions; phrase judgment calls as decision rules. Give a one-line reason for unusual constraints.
• State autonomy boundaries once: what agents may do without asking and what requires confirmation. Do not scatter "ask first" across sections.
• Include success criteria and verification gates for common workflows.
• If a rule must run at a fixed lifecycle point, document or create a hook instead of relying only on an instruction file.
• For task-specific procedures that do not need to load every session, create a skill instead of adding more content to AGENTS.md or CLAUDE.md.
• Avoid asking agents to reveal chain-of-thought or echo internal reasoning. Ask for conclusions, checks, evidence, diffs, or concise rationale instead.
• Avoid instructions that conflict with host-agent system messages, sandbox rules, approval policies, or tool availability.
• Do not copy large sections from README.md; link or summarize only what an agent needs during code work.
• Never include API keys, credentials, connection strings, or vulnerability details; the file ships with the repo and enters every prompt.

Output Contract

After initializing, report:

• files created or changed;
• whether CLAUDE.md imports @AGENTS.md, is a symlink, or is separate;
• the main commands and conventions captured;
• validation performed and any commands not run.

Examples

Good:

• "Initialize AGENTS.md and CLAUDE.md for this repo."
• "Create project memory for Codex and Claude Code."
• "Analyze the project and generate a best-practice AGENTS.md."

Near miss:

• "Add this testing note to AGENTS.md." Use the maintenance skill.