菜单
Use when a diff needs PR-style review, or findings need atomic repair and same-scope re-review — Standards axis and Spec axis, kept separate.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when writing or designing code: name premise and blast radius, deepen modules at seams, TDD at agreed seams, diagnose before patching, and keep producer→consumer contracts honest.
Use when premises or options must be pressure-tested, a short plan/handoff is needed, or a large plan must be sliced — grill facts locally, ask decisions one at a time.
Use when a project needs a lean `.do-it/handbook/` for stable project truth plus `.do-it/worklog/` for daily or goal notes.
Use before any done, fixed, passing, ready, install, or merge claim, and when closing a branch — fresh evidence on this worktree; hooks vary by host, this skill is the checklist.
Use only when the user explicitly asks to enable, disable, inspect, report, or persist opt-in local do-it behavioral feedback.
Use when starting any non-trivial repo task: pick Light / Standard / Heavy tier, then self-select meaning skills — not a fixed pipeline.
| name | do-it-review |
| description | Use when a diff needs PR-style review, or findings need atomic repair and same-scope re-review — Standards axis and Spec axis, kept separate. |
Findings-first review, then atomic fix + re-review. Two axes stay separate so neither masks the other.
Leading words: Standards, Spec, smell, clean.
| Axis | Question |
|---|---|
| Standards | Does the code follow this repo’s conventions + a smell baseline? |
| Spec | Does the diff faithfully implement the originating ask / plan / acceptance? |
A change can pass one and fail the other — report both. Prefer parallel reviewer agents when intensity is adversarial so contexts do not pollute each other.
Smell baseline (judgement calls; repo docs override): Mysterious Name, Duplicated Code, Feature Envy, Data Clumps, Primitive Obsession, Repeated Switches, Shotgun Surgery, Divergent Change, Speculative Generality, Message Chains, Middle Man. Skip what tooling already enforces. Lens detail: ../references/review-lenses.md.
Heavy, interface-breaking, non-local risk, or parent asked. Light mechanical: quick inline check.
| Intensity | When |
|---|---|
| review-quick | Light+code; Standard local/low risk — parent inline |
| review-deep | Standard non-local; Heavy without high-risk lenses — one reviewer |
| review-adversarial | Heavy + interface/packages/security/migration — multi-lens / parallel axes |
../references/workflow-kernel.md § Finding Schema:
severity / location / issue / cause_class / required_fix / NOT_CHECKED
One complete batch, severity-ordered. Not clean while Blocking/Important remain.
A named reviewer, a marker, or a claimed review pass is not evidence. Choose depth from concrete failure modes and inspect the changed proof path; accept a finding only with code, diff, contract, or command evidence.
Scope frozen; both axes checked at required depth; evidence fresh; no open Blocking/Important; Opportunities deferred only when named.
NEEDS_CONTEXT when unreproducible or out of ownership; BLOCKED when the same
finding survives one targeted repair + verify. Do not invent other stop statuses
(STILL_OPEN is retired — see workflow-kernel).