Claude Code Build Kit — Productivity skill for Claude Code

claude-code-build-kit

An opinionated build workflow for [Claude Code](https://www.claude.com/product/claude-code) — the orchestrator, hooks, and discipline rules that make multi-commit features ship without drift.

This is the kit, not the kit's documentation. Every file in `claude/` is something currently running on the maintainer's machine. Read [ARCHITECTURE.md](ARCHITECTURE.md) for the system design and [INSTALL.md](INSTALL.md) for setup.

What you get

Drop this kit on top of an existing Claude Code install and you get:

1. /build — a single skill that runs an end-to-end feature build: risk scorecard → plan → Codex audit → atomic chunked commits → multi-reviewer Codex review per commit → QA → summary. A forced five-dimension risk scorecard routes every build to a fast / standard / full tier (no flag to skip it); standard and full tiers add a bounded adversarial layer — a per-chunk "definition of done" checklist locked before any code (Codex adds the edge cases), run against each chunk, plus one whole-system integration pass at QA. Reads a manifest at /plan/.build-state.json so a session can resume itself without re-asking for context. Push to main is opt-in — declare it in your repo's CLAUDE.md or AGENTS.md if you want auto-push on QA pass.
2. A 10-section CLAUDE.md — operating principles (think before coding, simplicity first, surgical changes, persistence, unbiased consults) loaded into every conversation.
3. 3 hooks that /build depends on at runtime: a Codex audit on every plan exit, a multi-reviewer Codex audit on every commit, and vibecop adjudication on every edit.
4. A handoff script that transitions a heavy session into a fresh one without losing context — drops a prompt file, opens a new Terminal tab, the new session picks up via a .zshrc cooperator.
5. Codex CLI defaults at gpt-5.5 + medium reasoning, set globally. Every script in the kit inherits them; none passes -m or -c overrides.

Har