Entry/Exit
- Entry: Review feedback (from a human partner or an external reviewer)
- Exit: Feedback response after processing item by item
- Points to: After processing is complete, return to
/buildFix - Preload: CANON.md
- Output path:
docs/features/YYYYMMDD-<name>/04-review.md(feedback response part) → build-workflow-execute (fix implementation)
When not to use
- There is no specific review feedback yet, only preparation to proactively review
- The feedback has already been verified and converted into clear plan/build tasks
- The user asks to re-review the artifact rather than process feedback already received
Iron Law
Process
Step 1: READ — Read completely
Read all feedback completely without any reaction. Do not make changes while reading. Record the number of feedback items.
Step 2: UNDERSTAND — Understand the intent
Restate the request of each piece of feedback in your own words. If you cannot restate it = you did not understand → ask for clarification. Ask separately about each unclear piece of feedback.
Step 3: VERIFY — Verify the factual basis
Verify the feedback against the codebase: Does the corresponding code exist? Does the described phenomenon really exist? Is the referenced context complete? If the facts do not hold → record evidence and prepare a technical response.
Step 4: EVALUATE — Evaluate reasonableness
Evaluate against existing patterns in the codebase. Consider execution constraints (performance, compatibility, migration cost) and implementation risk. Give a judgment: accept / rebut / partially accept.
Step 5: RESPOND — Technical response
Accept: explain the understanding and technical basis, not empty agreement. Rebut: facts + data + alternative. One response per piece of feedback.
Step 6: IMPLEMENT — Implement item by item
Arrange in implementation order, handle one at a time, and run tests after each item. If tests fail → roll back that change and re-evaluate.
Source distinction
trusted human partner: Implement directly after understanding. Steps 1-2 are still required. Ask questions about parts that are not understood.
External reviewer: Must pass the 5-point verification checklist: Are the facts accurate? Is the context complete? Are constraints enforced? YAGNI check? Is the implementation cost reasonable?
YAGNI check
When a reviewer suggests "properly implement" or "add abstraction", first grep to confirm real usage scenarios: 1 = do not abstract, 2 = judge by style, 3+ = abstraction is reasonable. No third usage scenario = do not abstract.
Rebuttal guide
When to rebut: There is technical evidence, there is quantified impact, or it violates existing codebase patterns with no sufficient justification.
How to rebut: Facts + data + alternative. Example: "This change adds ~200ms latency (baseline X ms); the alternative is Y, because [reason]."
How to correct a rebuttal: Directly say "my previous rebuttal does not hold, because...", make no excuses, and immediately switch to implementation mode.
Prohibited behaviors (red flag area)
| Prohibited wording | Why prohibited |
|---|---|
| "You're absolutely right!" | Empty agreement, no technical analysis |
| "Great point!" | Agreement without analysis |
| "Thanks for catching that!" | Thanks is not a technical response |
| "I'll fix that right away" | Committing without evaluating |
| Saying "okay" to all feedback | yes-machine mode |
| Silently accepting all suggestions | Giving up independent judgment |
Implementation order
- Clarify unclear items FIRST — ask about what you don't understand, don't guess
- Blocking issues — Critical level
- Simple fixes — Nits that can be handled quickly
- Complex fixes — Items that require design or refactoring
Common excuses
| Excuse | Reality | Consequence |
|---|---|---|
| "The reviewer is always right" | Blind trust is as dangerous as blind rejection. | Blindly accepting ~20% inapplicable feedback introduces new problems |
| "You can't rebut the reviewer" | A rebuttal backed by evidence is a contribution. | Suppressing dissent → review degrades into rubber-stamping |
| "Finish changing everything first, then talk" | Batch acceptance = giving up judgment. | Cannot locate which item introduced a new bug; rollback scope = everything |
| "There's too much feedback, just pick and choose what to change" | Read and evaluate every item; you must not skip reading. | Unread feedback may contain Critical issues |
| "The reviewer knows better than I do" | The reviewer sees a snapshot; you live in the codebase. | Blindly executing snapshot judgment overrides lived experience |
Violating the letter of the rule is violating its spirit. There is no gray area.
Handling verification failures
| Failure scenario | How to handle |
|---|---|
| Test fails after implementation | Roll back that change, re-evaluate, do not continue to the next item |
| After rebutting, find it does not hold | Directly admit the mistake, switch to implementation mode |
| The facts in external feedback do not hold | Record the evidence, provide a technical rebuttal |
| The intent of the feedback is unclear | Mark as "pending clarification", do not guess the intent |
| After batch changes, cannot locate the issue | Roll back everything, switch to implementing one by one |
Red flags
- Starting to change code before reading all feedback
- Saying "okay" to all feedback (yes-machine)
- Rebutting without technical evidence (purely subjective preference)
- Confusing "I disagree" with "you are wrong"
- Continuing to the next item after one fails (no rollback)
- Skipping tests and going straight to the next item
- Treating "thanks" as a technical response
- Guessing the intent of feedback without asking
Verification checklist
- Every item of feedback has been read
- Questions have been asked about what is not understood
- External feedback has passed the 5-point verification checklist
- Every decision is backed by technical evidence
- Implementation order is correct (clarify → blocking → simple → complex)
- Tests pass after each implementation
- Response recorded in
docs/features/<name>/04-review.md
Output template
## Review Feedback 响应
### 反馈来源
- 来源: human partner / 外部 reviewer
- 反馈条数: N 条
- 处理状态: 全部评估 / 部分待澄清
### 逐条响应
| # | 反馈摘要 | 评估结论 | 理由 | 实施状态 |
|---|---------|---------|------|---------|
| 1 | [摘要] | 采纳/反驳/部分 | [技术依据] | 已修复/已反驳 |
### 待澄清项
| # | 反馈摘要 | 不明确之处 | 状态 |
|---|---------|----------|------|
| 7 | [摘要] | [描述] | 待 reviewer 补充 |
### 实施验证
- 全部修改后测试: PASS
- 无回归: 确认