Agent onboarding

Purpose

Initialize a repository so that AI coding agents (Claude Code, Codex, Cursor, etc.) can operate safely, deterministically, and with minimal human supervision.

This skill generates:

• AGENTS.md (agent operational contract)
• docs/ directory with structured technical documentation

Inputs

• Repository to initialize.
• Optional: short human note describing the project and instructions (1–3 sentences)

Outputs

• AGENTS.md at repository root
• docs/ directory with standardized markdown files

Core Principles

• Do NOT guess. If information is missing, explicitly mark TODOs.
• Prefer correctness and safety over completeness.
• Optimize for agent execution, not human prose.
• Minimize exploration cost and failure modes.
• If the given repository follows a monorepo pattern, create a subdirectory under docs/ for each project so the documentation doesn’t get mixed together.
  • Example) docs/frontend/00_overview.md docs/frontend/adr docs/frontend/plan

---

Execution Steps

Step 1: Repository Inspection

• Detect:
  • Primary language(s)
  • Framework(s)
  • Package manager
  • Test framework
  • Deployment indicators (Dockerfile, CI configs, Vercel, etc.)
• Identify:
  • Entry points
  • Core business logic directories
  • Infrastructure-related files (DB, queues, storage)

If detection is ambiguous:

• Record ambiguity explicitly in docs/00_overview.md
• Do NOT assume defaults

---

Step 2: Confirm Ambiguities & Decisions With the Human

Request the user to provide the following information:

• Whether Git-related commands are allowed: use freely vs use only with explicit approval
• Language preferences:
  • Language for comments, documents and commit messages
  • Communication language: same as the user vs a specific language
• Any other items discovered in Step 1 that are ambiguous or seem to require questions

If the human does not answer:

• Do NOT guess.
• Pause and ask again as needed; if you must proceed, mark missing items as “Unknown / TODO” in docs (without embedding the unanswered questions), and keep AGENTS.md conservative (minimal privileges, no destructive ops).

---

Step 3: Generate AGENTS.md

First, copy AGENTS.md.example, then edit it based on the information gathered so far to create AGENTS.md.

AGENTS.md MUST include:

Repository Guidelines

Project Structure & Module Organization

Documentation

Coding Style & Naming Conventions

Commit Message & PR Guidelines

Commit Message Format

Atomic Commits

Agent Operation & Safety Rules

Language Policy

User Interaction

Hard rules:

• Keep AGENTS.md concise (target: 50–120 lines)
• Link to docs/ instead of duplicating explanations

---

Step 4: Generate docs/ Structure

If the docs/ directory already exists, generate all new files inside docs/agents/ instead of docs/. Otherwise, use docs/ as the target directory.

Create the following files if applicable:

• docs/00_overview.md
• docs/10_architecture.md
• docs/20_dev_env.md
• docs/30_commands.md
• docs/40_testing.md
• docs/50_deploy.md
• docs/60_observability.md
• docs/70_security.md
• docs/80_style_guide.md
• docs/90_troubleshooting.md

Rules:

• Omit files that are clearly irrelevant
• Never fabricate cloud providers, CI tools, or databases
• Prefer explicit “Unknown / TODO” over assumptions

---

Step 5: ADR Detection (Optional but Recommended)

If architectural decisions are inferred:

• Create docs/adr/
• Record assumptions as provisional ADRs
• Mark them as “UNVERIFIED”

---

Step 6: Validation Checklist

Before finalizing:

• All commands copy-paste runnable OR marked TODO
• No secrets included
• No speculative tech choices
• AGENTS.md links resolve correctly

---

Failure Handling

If the repository is:

• Empty
• Too small to infer structure
• Highly ambiguous

Then:

• Generate minimal AGENTS.md
• Generate docs/00_overview.md that clearly separates detected facts vs “Unknown / TODO” items (but do not include a question list; ask any needed questions during execution).

If, at any point, you encounter ambiguities or actions that appear potentially risky:

Then:

• always pause and request clarification from the user before proceeding.

---

Non-Goals

• Do not optimize for marketing or onboarding humans
• Do not refactor code
• Do not introduce new dependencies
• Do not modify runtime behavior

---

Success Criteria

A coding agent can:

• Identify where to make changes
• Execute tests confidently
• Avoid destructive or unsafe operations
• Complete scoped tasks without human correction