AI 工程从零到一（中文）

Given a builder agent already producing workbench artifacts, stand up a reviewer that reads them and writes structured reports.

Produce:

1. agents/reviewer.md with the reviewer system prompt: read-only access, five-dimension rubric, must cite the artifact path for each score.
2. tools/reviewer.py that loads ReviewerInputs from the workbench and runs the LLM scorer per dimension.
3. outputs/review/<task_id>.json as the canonical review report path.
4. docs/reviewer-rubric.md listing the five dimensions, the question each one answers, and the 0-1-2 anchor descriptions.
5. CI step that posts the review report as a PR comment whenever a builder task closes.

Hard rejects:

• A reviewer with write access to the diff. The gap between builder and reviewer is the whole signal; collapsing it destroys reliability.
• A rubric without anchor descriptions per score. "Score from 0 to 2" without anchors collapses to vibes.
• Review reports that omit citations. Every score must point at a file or trace entry.
• Sharing the builder's system prompt. Same model is fine; same prompt is not.

Refusal rules:

• If the builder produces no verification report, refuse to run the reviewer. Acceptance must hold before judgment is worth asking for.
• If the project has fewer than three closed tasks, refuse to claim the rubric is calibrated. Save the first reports as the calibration set.
• If the reviewer is asked to score below a minimum confidence, refuse and surface the uncertain dimension to a human.

Output structure:

<repo>/
├── agents/reviewer.md
├── tools/reviewer.py
├── outputs/review/
│   └── <task_id>.json
├── docs/reviewer-rubric.md
└── .github/workflows/review.yml

End with "what to read next" pointing to:

• Lesson 40 for the handoff packet that combines verification + review.
• Lesson 41 for the real-style task that exercises builder/reviewer separation end to end.
• Lesson 05 (Self-Refine and CRITIC) for the single-agent self-review baseline this lesson improves on.