/review-step: Self-review (TDD step 7)

SkillDev tools

Step 7 of the Weft TDD workflow. Use after /e2e: self-reviews the diff via the code-review skill (medium effort; high when packages/core or packages/dom public API is touched) plus a spec-conformance check, then fixes or explicitly rejects every finding and re-validates until clean. Hard gate: no commit until the review is clean. Named review-step to avoid the built-in /review (GitHub PR review).

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the /review-step: Self-review (TDD step 7) skill

What this skill tells your AI

The instructions your AI receives, as published by stefvw93/weft in .claude/skills/review-step/SKILL.md and read by ahel’s review.

Review the full diff for correctness and spec conformance; loop until clean. Commits are blocked until this step passes.

When to run

  • Previous step: /e2e (browser tests green, or explicit skip recorded).
  • Next step: /document.
  • Gate (exit): a review pass with zero unaddressed findings. No commit until clean.

Procedure

  1. Pick the effort level.

    • Default: run the code-review skill at medium effort (fewer, high-confidence findings, fast loop).
    • Escalate to high when the diff touches packages/core or the public API surface of packages/dom (exports reachable from @weftui/dom, ./client, or ./server).
  2. Run the code-review skill on the working diff at the chosen level, pinned to Sonnet 5 (model: sonnet) regardless of the session's currently selected model. This step fans out several sub-agents; pinning keeps token cost predictable instead of inheriting a pricier model.

  3. Spec-conformance check (always, regardless of effort level). Independently of the skill's findings, verify:

    • The implementation satisfies every acceptance criterion in the co-located specs.md.
    • The implemented exports still match the mock surface exactly (signatures, error union): no drift since /implement's parity pass.
    • specs.md skip records (type-tests: not applicable, e2e: not applicable) are present where those steps were skipped.

    Any conformance gap is a finding, same weight as a bug.

  4. Disposition every finding (no silent drops). Each finding is either:

    • Fixed, or
    • Explicitly rejected with a stated reason, reported to the user.

    If a finding reveals spec/mock drift: pause rule. Stop, update specs.md and the mock surface (and affected tests) first, then resume.

  5. Re-validate and re-review loop. After applying any fixes:

    1. vp run check and vp run test (and vp run test:browser if browser-covered code changed): must be green.
    2. Re-run the review on the updated diff.
    3. Repeat until a review pass comes back clean (no findings, or only findings already explicitly rejected with reasons).
  6. Report the final state: findings fixed, findings rejected + reasons, number of loop iterations.

  7. Hand off. Next step is /document.

Rules

  • This step reviews the working diff pre-commit. Reviewing a GitHub PR is the built-in /review, a different tool.
  • A clean review is a hard gate: no commit, no PR, until it passes.

Signals

GitHub stars
44
Forks
1
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
review-step
Source
github.com/stefvw93/weft