| name | Agent Bootstrap |
| description | Use when setting up or repairing repo agent instructions for Claude Code, Codex, or both. Proposes durable AGENTS.md, CLAUDE.md, and project memory changes before writing unless setup was explicitly requested. |
| allowed-tools | Read, Bash, Edit, MultiEdit, Write |
Agent Bootstrap
Use this skill to add durable project instructions without overwriting repo intent.
When To Use
- User asks to bootstrap Claude Code, Codex, or AI agent workflow.
- Repo lacks useful
AGENTS.md / CLAUDE.md and the user wants setup or repair.
- Existing instructions are stale, duplicated, unsafe, or too large.
- Long project needs small boot files plus memory docs.
Do not use for normal coding unless instructions are part of the task.
Safety Rules
- Read existing
AGENTS.md, CLAUDE.md, README, package scripts, and relevant docs first.
- Never overwrite existing instructions blindly.
- Preserve project-specific rules; append guarded updates when possible.
- Keep boot files short: commands, safety rules, memory location, verification expectations.
- Put detailed context in
docs/agent-memory/, not in boot files.
- Do not add secrets, credentials, private paths, or one-project assumptions to reusable skills.
- If boot files or
docs/agent-memory/ are missing, do not auto-create them. Propose the files, summarize each in one line, and wait for explicit approval.
- Exception: create directly only when the user explicitly asked for project setup, agent setup, memory setup, or said to set up the repo.
Boot File Shape
Include only:
- project identity and critical flows
- verified setup/test/build commands
- destructive-action and secret rules
- where memory docs live
- when to update memory
- final verification expectations
Workflow
- Detect target: Claude Code (
CLAUDE.md), Codex (AGENTS.md), or both.
- Inspect existing files and commands.
- Patch existing boot files when requested or clearly part of the task.
- If files are missing, propose creating compact boot files and minimal memory index unless setup was explicitly requested.
- Run diff review. Confirm no secrets or huge prompt dumps.
Output
Report files created/changed, preserved existing rules, and any assumptions not verified.