Bootstrap CLAUDE.md

Author a new CLAUDE.md that orients a future agent to the current working scope: what it is, how it's wired, what's load-bearing, and the conventions a contributor must respect. Optimized for both human skim and LLM ingestion.

Output modes

Before drafting, ask the user which mode to use:

• Flat (default) — single CLAUDE.md containing every section. Best for leaf packages, small apps, anything that fits comfortably in ~150 lines.
• Progressive disclosure — slim CLAUDE.md orientation index that lists must-know rules + a table of links into context/*.md files for everything else. Best for large apps with multiple subsystems, where loading the whole file every turn is wasteful.

Quote the two options to the user and wait for their pick. If they don't choose, default to flat. Skip this prompt only if the user already named the mode in their request (e.g. "bootstrap a progressive claude.md").

Scope rules

• Target path: <cwd>/CLAUDE.md unless the user names a different directory. Resolve to an absolute path before writing.
• If the file already exists: do NOT silently overwrite. Stop and ask the user how to proceed (see step 1). The user picks: route to update-claude-md to refresh in place, or delete the existing file and start fresh from this skill. Only proceed once they've chosen.
• Never touch ~/.claude/CLAUDE.md or sibling packages' CLAUDE.md files.
• Inherit, don't duplicate: walk up to the repo root and read every ancestor CLAUDE.md. Anything already covered there (global naming rules, lint config, monorepo tooling) should NOT be repeated — only note local deviations.