/ai-os-init

name: ai-os-init
description: >-
  Scaffolds an "AI Operating System" project structure into any project,
  non-destructively — adds only missing pieces, never removes or overwrites
  anything that already exists. Creates: CLAUDE.md (memory layer),
  docs/architecture.md + docs/decisions/ + docs/runbooks/ (documentation),
  tools/scripts/ and tools/prompts/ (utilities), and .claude/ with two example
  skills (new-adr, clean-tests), six hooks (guard-secrets, branch-guard,
  auto-format, typecheck, n+1-guard, audit-log), and a read-only docs-auditor
  subagent — all wired via .claude/settings.json. Requires only Python 3 and
  bash; no external dependencies. Use when asked to "init Claude Code project
  structure", "scaffold the AI-OS layers", "add CLAUDE.md hooks subagents to
  this project", "set up the Claude Code layers", "initialize ai-os", "add the
  AI operating system structure", "scaffold this project for Claude Code", or
  "add skills hooks agents here". Safe to re-run — idempotent. Not for
  non–Claude-Code agents (it scaffolds Claude Code's .claude/ layer) and not a
  code generator — it sets up project structure, not application code.

ai-os-init

Scaffolds the five-layer AI Operating System structure into any project directory. **Non-destructive by design**: existing files are never touched, hook entries are never duplicated, source trees are never imposed.

**No dependencies** — needs only Python 3 and bash. The hooks that call formatters or type checkers (ruff, prettier, tsc, mypy, …) skip silently when those tools aren't installed, and every hook allows the action (fails open) if `python3` itself is missing, so nothing ever breaks.

The Five Layers It Wires Up

| # | Layer | What gets created | |---|---|---| | 01 | **Memory** | `CLAUDE.md` — always-loaded project map, conventions, links | | 02 | **Knowledge** | `.claude/skills/new-adr/` + `.claude/skills/clean-tests/` — on-demand skills | | 03 | **Guardrails** | `.claude/hooks/` — six PreToolUse/PostToolUse hooks (see below) | | 04 | **Delegation** | `.claude/agents/docs-auditor.md` — read-only docs drift subagent | | 05 | **Docs** | `docs/` architecture + decisions (ADR log) + runbooks |

Hooks installed (wired into `.claude/settings.json`)

| Hook | Event | What it does | |---|---|---| | `guard-secrets.sh` | PreToolUse (Edit/Write) | **Blocks** writes to secret-looking files (.env, keys, credentials); allows `.env.example`-style templates | | `branch-guard.sh` | PreToolUse (Bash) | **Blocks** force-push (`--force-with-lease` allowed) and commit/push to protected branches; runs local checks before push | | `auto-format.sh` | PostToolUse (Edit/Write) | Formats the edited file (ruff/black/prettier/gofmt/rustfmt), best-effort | | `typecheck.sh` | PostToolUse (Edit/Write) | Feeds type errors (tsc/mypy) back to Claude if the project is configured for them | | `n+1-guard.sh` | PostToolUse (Edit/Write) | Warns Claude about likely N+1 query patterns (DB call inside a loop) | | `audit-log.sh` | PostToolUse (Bash) | Silently appends every bash command to `.claude/command-audit.log` |

How they talk to Claude (per the [hooks reference](https://code.claude.com/docs/en/hooks)): blocking hooks exit `2` with the reason on **stderr**; advisory PostToolUse hooks print JSON `hookSpecificOutput.additionalContext`, because plain stdout from a PostToolUse hook only reaches the debug log. Commands are registered as `bash "${CLAUDE_PROJECT_DIR}/.claude/hooks/<name>.sh"` so they keep working after Claude `cd`s into a subdirectory.

> Two of these change git/agent behaviour out of the box: **`branch-guard.sh`** > will block commits on `main/master/develop/staging/production` and force > pushes, and **`audit-log.sh`** logs every command. Tell your team these are > active. The audit log can capture secrets typed into commands — keep > `.claude/command-audit.log` in `.gitignore` (the scaffold reminds you if it > isn't).

How to Run

`$SKILL_DIR` is the folder this `SKILL.md` lives in — `${CLAUDE_PLUGIN_ROOT}/skills/ai-os-init` for a plugin install, or the skill's own directory for a manual / `npx skills` install.

python3 "$SKILL_DIR/scaffold.py" --dry-run   # preview: lists what would be created/merged, writes nothing
python3 "$SKILL_DIR/scaffold.py"             # scaffold the current directory
python3 "$SKILL_DIR/scaffold.py" /path/to/project

If the user seems unsure, or the project already has a `.claude/` setup, run `--dry-run` first and show them the list before writing anything. Afterwards, report the created / merged / skipped summary to the user, and mention that new hooks show up in `/hooks` (or after a restart).

Non-Destructive Contract

• **File exists →** skip it, report it as "already present"
• **File missing →** create it from the template
• **`.claude/settings.json` →** load and merge; append each hook entry **only

if** no existing hook already runs that script (matched on `.claude/hooks/<name>.sh`, so entries from older versions aren't duplicated) — then write back preserving every existing key and hook. A new file also gets the `$schema` line for editor autocomplete. If the file can't be parsed, it is left untouched and the merge is skipped.

• **Dangling symlink at a target path →** treated as present; never written

through

• **`CLAUDE.md` exists →** leave it byte-identical; note which recommended

sections are absent so the user can fill them in

• **No `src/` imposed** — if the project has `src/`, `lib/`, or `app/` but

no nested `CLAUDE.md`, suggest adding one; never create source dirs

What to Do After Scaffolding

After running, guide the user through: 1. Fill in `CLAUDE.md` — project name, conventions, key commands 2. Fill in `docs/architecture.md` — tech stack, components, data flow 3. Try `new-adr`: say "create an ADR about [decision]" 4. Try `clean-tests`: say "clean up the tests" 5. Try `docs-auditor`: say "audit the docs" 6. `guard-secrets` and `branch-guard` are activ