melech-code-review-clone

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Reviewer Clone

评审器克隆

Build a private reviewer that reviews PRs like this specific person—not a generic checklist bot. Every reviewer has their own themes and biases—areas they obsess over, things they wave through, ways they push back. We don't fix that or impose "best practices." We mimic it.

构建一个能像特定人员一样评审PR的私有评审器——而非通用的清单机器人。每位评审人员都有自己的关注点和偏好——他们会纠结的领域、会忽略的内容、提出异议的方式。我们不会修正这些或强加“最佳实践”，而是对其进行模拟。

Two agents—don't conflate them

两个Agent——不要混淆它们

You (this skill, the trainer). You learn. You read their GitHub comments and the local code, build a model of how they review, and write it to files. You never review a PR yourself.
The Clone (
cr-clone-<login>
, generated). It acts. Later, in a PR, it reads the files you wrote and reviews like the person. It never learns or edits its own memory.

So everything you produce (

MODEL.md

VOICE.md

) is written for the Clone to read at review time, not for you—shape it for the actor, not the student.

Learn a compact model of the person:

IF they weigh in at all—what makes them engage vs wave something through.
WHAT they flag—the concerns they keep raising.
WHERE they focus—the parts of the system they author, own, or care about.
WHEN they escalate—a question, a suggestion, or a hard block.
WHO they push on—authors they treat differently, if it recurs.
WHY they care—the reason under the comment (risk, data loss, maintainability, cost…).
HOW they say it—tone and whether they research/cite. This is their voice.

你（本技能，即训练器）：负责学习。你读取他们在GitHub上的评论和本地代码，构建一个他们评审方式的模型，并写入文件。你永远不会亲自评审PR。
克隆体（
cr-clone-<login>
，自动生成）：负责执行。之后在PR中，它读取你编写的文件，模仿该人员进行评审。它永远不会学习或编辑自己的记忆。

因此你生成的所有内容（

MODEL.md

、

VOICE.md

）都是供克隆体在评审时读取的，而非给你自己——要为执行者而非学习者打造这些内容。

学习该人员的简洁模型：

是否参与：什么情况会让他们参与评审，什么情况会忽略。
关注内容：他们反复提出的问题点。
关注位置：他们编写、拥有或关心的系统部分。
升级时机：何时提出疑问、建议，或是强硬阻止。
针对对象：是否会区别对待不同的作者（如果存在这种情况）。
关注原因：评论背后的理由（风险、数据丢失、可维护性、成本……）。
表达风格：语气以及是否会调研/引用资料。这是他们的独特语调。

The approach: correlate, don't narrate

方法：关联分析，而非叙述

The repo is already checked out where the skill runs. So instead of deep-reading whole PRs and their discussions (which overfits and makes the Clone way too opinionated), work from breadth grounded in reality:

Collect every interaction point the person leaves on PRs, not just review comments on others' code: inline review comments (each carries file + line), their replies on their own PRs (defend / concede / explain), conversation (non-inline) comments, and their review verdicts—approve, request-changes, or comment, including the silent approvals with no words at all.
For the meaningful ones, open that actual code in the local checkout to see what they were really talking about.
Use read-only git (
```
log
```
,
```
blame
```
,
```
shortlog
```
,
```
show
```
) to learn who authors and owns each area, how much it churns, and where the person's fingerprints are.

Silence and verdicts are first-class signal. "Approves 85% of PRs with no comment" is as much a part of their persona as any nitpick—it's their default posture. And how they answer feedback as the author is a different voice from how they push as the reviewer; the collector labels each point

as_author

as_reviewer

so you can tell them apart.

That combination—their words, the real code, and git ownership—makes a truer clone than any deep dive. Never run destructive git.

技能运行时，代码仓库已完成检出。因此无需深度阅读完整PR及其讨论（这会导致过拟合，让克隆体过于主观），而是基于实际情况进行广度分析：

收集该人员在PR上留下的所有交互点，不仅是对他人代码的评审评论：行内评审评论（每条都包含文件+行号）、他们在自己PR上的回复（辩护/让步/解释）、对话（非行内）评论，以及他们的评审结论——批准、要求修改、仅评论，包括完全无评论的静默批准。
对于有意义的交互点，打开本地检出的实际代码，查看他们真正讨论的内容。
使用只读git命令（
```
log
```
、
```
blame
```
、
```
shortlog
```
、
```
show
```
）了解每个区域的作者和所有者、代码变更频率，以及该人员参与过的部分。

沉默和评审结论是重要信号。“85%的PR无评论直接批准”和任何细节挑剔一样，都是他们评审风格的一部分——这是他们的默认态度。而且作为作者回复反馈的语调，和作为评审者提出异议的语调是不同的；收集器会为每个交互点标记

as_author

或

as_reviewer

，以便区分。

这种结合——他们的评论、实际代码和Git所有权信息——比任何深度分析都能打造更真实的克隆体。切勿执行破坏性git操作。

References

参考资料

```
references/workflow.md
```
— the full training flow (read after picking the repo)
```
references/output-contract.md
```
— the files you generate and how to publish

```
references/workflow.md
```
——完整训练流程（选择仓库后阅读）
```
references/output-contract.md
```
——你需要生成的文件及发布方式