shipwright

Hand it a requirement, get back a PR with proof.

Shipwright is a Claude Code skill that drives a feature or bugfix from one sentence to a shipped pull request — through a disciplined, end-to-end pipeline that brainstorms, plans, builds with TDD, reviews, runs browser QA against the real API, writes an evidence-rich finalize report, and opens the PR. It runs interactively, or fully autonomously from a single requirement after one approval.

No half-baked diffs. No skipped tests. No "works on my machine." Just high-quality, evidence-backed pull requests.

It's an orchestrator: each phase invokes a battle-tested skill and waits for it to finish. The discipline comes from Andrej Karpathy's rules for LLM coding — don't assume, simplicity first, surgical changes, goal-driven — propagated into every phase and every subagent prompt.

🏗️ The Shipwright pipeline

Shipwright doesn't just write code — it runs a strict engineering lifecycle, every time:

[🧠 Brainstorm] ➔ [📋 Plan] ➔ [🧪 TDD Build] ➔ [🔍 Review] ➔ [🌐 Browser+API QA] ➔ [📊 Finalize] ➔ [🚀 PR]

• 0 🔍 Preflight — verifies dependencies before doing anything (gstack must be installed; superpowers, karpathy, and frontend-design are bundled).
• 1 🧠 Brainstorm — deconstructs the requirement, maps edge cases, and agrees on an approach → an approved spec.
• 2 📋 Plan — generates a detailed, step-by-step implementation plan before touching files (auto-reviewed via /autoplan when running autonomously).
• 3 🧪 Build (TDD) — writes failing tests first, then the minimal clean code to pass them, with Karpathy discipline injected into every subagent prompt and frontend-design for UI work.
• 4 🔍 Review — self-reviews the diff for correctness, security, and style (+ /review, /codex for an adversarial second opinion).
• 5 🌐 Browser + API QA — drives a headless browser to verify the UI and asserts the real API calls returned 2xx via browse network (no blind mocking) — plus /design-review.
• 6 📊 Finalize + 🚀 Ship — produces a committed report (screenshots + per-stage cost table), passes a soft gate, and opens a structured, review-ready pull request.

requirement
    │
    ▼
[0] Preflight ──── verify gstack is installed (superpowers + karpathy + frontend-design are bundled)
    │
    ▼
[1] Brainstorm ─── shipwright:brainstorming → approved spec          ◀─┐
    │              └─▶ Autonomy gate: "run autonomously?" + QA creds   │
    ▼                                                                  │ re-scope
[2] Plan ───────── shipwright:writing-plans (+ /autoplan when auto)  ◀─┤
    │                                                                  │ re-plan / re-design
    ▼                                                                  │
[3] Build ──────── shipwright:subagent-driven-development · TDD · shipwright:frontend-design for UI
    │              karpathy discipline embedded in every subagent prompt
    ▼                                                                  │
[4] Review ─────── shipwright:requesting-code-review + receiving-code-review (+ /review, /codex)
    │                                                                  │
    ▼                                                                  │
[5] QA ─────────── gstack /qa · browse network (API 2xx) · console --errors · /design-review
    │              └─ finding bigger than a local fix? escalate ───────┘
    │                 (interactive: confirm first · autonomous: auto, scope-change stops)
    ▼
[6] Finalize ───── report + screenshots + per-stage cost table → soft gate → gstack /ship → PR
    │
    ▼
   PR with proof