atl init

Scaffold a lean starter CLAUDE.md for the tier you choose — only if one doesn't already exist, so your own CLAUDE.md is never overwritten.

CLAUDE.md is the file Claude Code auto-loads as project (and global) instructions. ATL ships a starter shape for each tier so you don't begin from a blank file; you fill in the parts marked for you, and the /brainstorm + /drain skills maintain their own marker blocks inside the project file over time.

Usage

atl init                 # a project-root CLAUDE.md (default)
atl init --project       # the same, explicit
atl init --global        # your personal ~/.claude/CLAUDE.md persona
atl init --monorepo      # a lean ~30-line orientation file

The three flags are mutually exclusive. atl install also drops the project starter automatically when a project has no CLAUDE.md (see What it does), so you usually only run atl init by hand for the global persona or a monorepo orientation file.

Tiers

Flag	Target path	Shape
--project (default)	<project>/CLAUDE.md	Hybrid: ATL-managed marker blocks (active brainstorms, knowledge index — maintained by /brainstorm + /drain) plus user-owned, evidence-fillable facts (stack, commands, conventions) and an optional skill-routing table. Soft budget ≤ ~60 lines.
--global	~/.claude/CLAUDE.md	Pure user persona — how you want Claude to work everywhere. ATL manages nothing here. Soft budget ≤ ~80 lines.
--monorepo	<repo>/CLAUDE.md	The project shape specialized + lean: a layout table and conventions as pointers, not inlined content. Soft budget ~30 lines.

The tiers, their budgets, and the managed-vs-owned ownership model are explained in full on the Claude Code conventions page.

What it does

atl init:

1. Resolves the target path for the chosen tier (global → ~/.claude/CLAUDE.md; project / monorepo → the project root's CLAUDE.md).
2. If a CLAUDE.md already exists there, it does nothing — your file is user-owned and never overwritten.
3. Otherwise it writes the tier's starter skeleton (filling the project / repo name) and prints the path it created.

For the project and monorepo tiers, atl init also drops empty .atl/backlog.md + .atl/tasks.md skeletons alongside the CLAUDE.md — the two decision-state files the /brainstorm skill keeps current (see Backlog & tasks). Each is written only if absent, so your own backlog.md / tasks.md is never overwritten. The global tier is skipped — it has no project .atl/.

atl install runs the same project-tier scaffold as a best-effort step: when you install a team into a project that has no CLAUDE.md, ATL drops the project starter (and the .atl/backlog.md + .atl/tasks.md skeletons) so the /brainstorm and /drain blocks have a home. It is only-if-absent and never fails the install.

Idempotency — safe to re-run

Re-running atl init (or atl install) when a CLAUDE.md is already present is a no-op — it reports that the file exists and leaves it untouched. There is no --force; replacing an existing CLAUDE.md is a deliberate manual act, not something a scaffold should do.

Related

• Claude Code conventions — the three tiers, token budgets, ownership model, and the marker blocks the project file carries
• Backlog & tasks — the .atl/backlog.md + .atl/tasks.md decision-state files this scaffold drops for the project / monorepo tiers
• atl install — installs a team and drops the project starter if absent
• /brainstorm — maintains the <!-- brainstorm:active --> block in the project CLAUDE.md
• /drain — maintains the <!-- wiki:index --> knowledge map block