from-scratch-bootstrap

The first session in a brand-new project shapes everything that follows. This skill enforces the nine phases of docs/graph/protocols/from-scratch.md so the first day produces a project an agent can pick up cold on day two.

When to apply this skill

• Repo is empty or only contains LICENSE / placeholder README.
• No docs/graph/plans/grill.md exists yet.
• The user just said "let's start a new project to do X".
• A mkdir for a new project is about to happen.

The phases live in the protocol; the honesty lives here

The nine-phase sequence and each phase's orchestration — what it produces, which sub-protocol drives it, its exit conditions — are owned by docs/graph/protocols/from-scratch.md (from-scratch.phases). This skill does not restate them; it is the discipline that keeps the first day from cutting the corners the sequence alone cannot prevent:

• Bootstrapping is inherently T3 (kernel §0). The full funnel is the proportional response, not ceremony — never discount it because the goal "sounds clear"; the brainstorm exists to surface the constraints the user did not state.
• The skeleton phase is done only when the host tool actually loads the machinery (agents, protocols, skills) from its own directory — not when the files merely exist. Claude Code → .claude/ + CLAUDE.md; Prime Agent → .prime/agent/ + AGENTS.md; opencode → .opencode/ + AGENTS.md; Codex → .codex/ + config entries; Copilot → .github/ + copilot-instructions.md.
• The stack is chosen from verified research, never memory — version, maintenance signal, and license confirmed and wikified before it is committed to (research routinely invalidates a remembered option).
• The verification baseline passes on a clean checkout before any feature code — a test framework configured and a smoke test green come first; a slice without a passing gate is a draft, not a slice.
• Specs before tests before code, always — the specify phase authors the spec, not the implementation.

Each rule is a way the first day silently goes wrong; the catalog below names the rest.

Common failure modes

• Skipping the brainstorm. "It's just a CLI" — wrong; the brainstorm uncovers constraints.
• Picking the stack from memory. Verify version, maintenance, license; wikify before committing.
• Writing feature code before gates pass. A slice without a passing gate is a draft, not a slice.
• Skipping ADR-0001. Future agents will ask "why this design" and have no answer.
• Skipping SPEC-0001. Future agents won't know what the slice was supposed to do, only what it does.
• Writing code in Phase 7 instead of the spec. Specs first; tests second; code third. Always.
• Phase 6 "we'll add tests later". A bootstrap without a test framework configured is incomplete.

Reference files

• docs/graph/protocols/from-scratch.md — the protocol that drives this.
• docs/graph/skills/brainstorm-socratic.md — Phase 1.
• docs/graph/skills/grill-planner.md — Phase 3.
• docs/graph/skills/research-and-ingest.md — Phase 4.
• docs/graph/skills/library-wiki.md — Phase 4.
• docs/graph/skills/adr-writer.md — Phase 5.
• docs/graph/skills/spec-author.md — Phase 7.
• docs/graph/protocols/test-first.md — Phase 8.
• install.sh — drops the seed system into the project per host tool.