← ClaudeAtlas

create-great-harnesslisted

Use when init finds an app with no CLAUDE.md and no .claude/ directory at all and the user agrees to bootstrap one, or when the user directly asks to set up a minimal agent system for an app that doesn't have one yet.
wayne930242/straw-boss · ★ 3 · AI & Automation · score 76
Install: claude install-skill wayne930242/straw-boss
## Overview A lightweight agent-system bootstrap for an app that has none: a short `CLAUDE.md` (Role/Scope/Standards, non-obvious content only), one verified guard hook for the single most common destructive footgun, and one rule that pins future skill-authoring in this app to the current official spec rather than stale training data. Not a scaffold of empty `.claude/skills/` directories or a full rules library — those earn their place only once there's real content to put in them; the skill-authoring rule is the one exception, because it costs nothing to keep current (a live fetch, not hand-maintained prose) and every app that ever grows a `.claude/skills/` needs it from the first skill written, not retrofitted after an off-spec one already exists. Runs either inline (a user asked for this directly) or as a dispatched agent rooted in `<app-dir>` (`init`'s own bootstrap step dispatches it via `dispatching-work`). The dispatch instruction, when present, states that scope was already confirmed and how to report completion — Task 1 and Task 6 both branch on whether that's the case. ## Task 1: Confirm scope before writing anything State exactly what will be created: `<app-dir>/CLAUDE.md`, one guard hook in `<app-dir>/.claude/settings.json`, and one rule in `<app-dir>/.claude/rules/`. - **Invoked directly by a user in this session:** get explicit confirmation before proceeding — this writes into the app's own checkout, not just plugin state. - **Invoked as a dispatched agent: