Simple Claude Code review skill wins six-month test
Juejin AI summarized developer mizoko’s six-month comparison of three reusable Claude Code pull-request review skills. Among 1,306 included comments, the basic review-pr skill had a 68.9% repair rate, ahead of nitpick-pr at 66.7% and the more complex deep-review at 59.2%, while all three had roughly 10% incorrect comments. The observational test covered one mobile app and one backend, excluded some unresolved pull requests, and could not attribute 36% of included comments, so the results are not broadly conclusive.
The comparison provides practical evidence that a simpler Claude Code PR-review workflow can produce a higher observed repair rate than a design using many parallel review perspectives, while also showing that comment volume, correctness, severity, and missed defects should be evaluated together.
2026-11-15,开发者 mizoko 在 dev.to 发表了《I Tested Three Claude Code Review Skills for Six Months. The Simplest One Won.》。过去半年里,他为一个移动端 App 和一个后端写了三个审查 PR 的 skill(skill 是 Claude Code 里可复用的指令包),攒下上千条审查评论并逐一核对。
在看最终结论之前,先把文中用到的统计范围和几个核心指标怎么算说清楚(作者在统计前先定死定义,中途不改):
- 统计范围:只纳入已合并、且 PR 作者回复过评论的 PR。
- 修复率:已修复的评论数 ÷ 这个 skill 发出的评论数。「已修复」指代码确实因为这条评论改了,作者根据之后的回复和提交判断。例如,
review-pr发了 354 条评论,其中 244 条已修复,修复率就是 68.9%。 - 错误评论比例:说错了的评论占这个 skill 全部评论的比例。作者没有给「错误」单独下定义;文中能看到的例子是 8 月那几条:指向的代码已经被更早的提交修掉、把有意为之的行为当成了问题、依据了一个错误的先例后来撤回。
- 每个 PR 评论数:平均每个 PR 收到这个 skill 的几条评论。
- 「正确但优先级低」与「超出这个 PR 的范围」:作者把没修的评论按原因拆开后得到的两类——评论本身没说错,但不急,或者不属于这个 PR 要解决的事,所以没改。比例按这个 skill 的全部评论算。