implement

Build a spec, PRD, issue, or clear feature request end-to-end with TDD, then run one code-review pass at the end across the whole change.

What it does

The defining rule: code review runs a single time over the complete diff, not after every phase. Per-phase reviews interrupt flow and re-flag churn that later phases delete; one review at the end sees the real diff.

Before writing code it reads the spec (and suggests brainstorm if scope is fuzzy), skims the surrounding code to match its patterns, and detects the task runner to prefer over calling tooling directly.

Build loop (per phase)

Work in small phases. Each phase:

1. Test first at a real integration point (TDD) — a branch, loop, parser, or money/security path, not every trivial line.
2. Minimum code to pass. No speculative abstractions.
3. Typecheck + run touched tests after each change. Keep the tree green as you go — a continuous safety net separate from the final review.
4. Move to the next phase. Do not spawn review agents here.

Finish (once, at the end)

1. Run the full suite + typecheck + lint, and fix failures.
2. Now run a single review pass over the entire diff — parallel lenses, confidence-gated. It reports; you act.
3. Triage findings: fix real issues, note deliberate simplifications, discard noise, then re-verify.
4. Commit with a conventional-commit message. To push and open a PR, hand off to ship.

When to use it

When scope is clear and it's time to build. Pairs with brainstorm before and ship after.