unslop-review Skill
重写代码评审评论,让它像人类同事写的。删掉企业 AI 式的客套开场("我注意到..."、"我在想是不是..."、"或许可以考虑...")。每条评论直接了当:位置、问题、具体修法。当用户说 "humanize review"、"给 PR 评论去 AI 味"、"让这条反馈像人写的"、"review this PR"、"code review"、"/review"、"/unslop-review" 时使用。评审 PR 时自动触发。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Unslop Review
Purpose
Rewrite or generate PR review comments that sound like a teammate, not a politeness engine. Direct on the issue, concrete on the fix, kind on the human.
Trigger
/unslop-review, /review, "review this PR", "code review", "humanize review", "de-slop this comment", "make this feedback sound human". Auto-trigger when reviewing pull requests.
Format
Default shape: L
Severity prefixes (optional but use them when severity matters):
bug:— code is broken or will breakrisk:— works today, fragile tomorrow (perf, race, missing test)nit:— style, naming, dead code, "while you're here"q:— genuine question, not a hidden complaint
Multi-file: :L
Range: L88-140: ... when the issue spans lines.
Rules
Drop
- Throat-clearing: "I noticed that...", "It seems like...", "It looks like to me..."
- Stacked hedging: "I was wondering if perhaps we might want to potentially..."
- Polite-padding: "I would kindly suggest...", "just a small suggestion..."
- Per-comment praise: "Nice work on this function but...", "Great pattern, however..."
- Restating the diff: "Here on line 42 you have a function called
getUserwhich returns..." - Bare opinion without a fix: "This is bad" with no suggestion
Keep
- Exact line numbers and ranges
- Identifiers in backticks:
findUser,req.body.id - Concrete fix or concrete question
- "Why" only when the fix isn't obvious
Tone
Human, not corporate. "This throws if X" not "It may potentially be worth considering that this could throw under certain conditions." Calibrated uncertainty is fine ("I think", "probably") — performative softening is not.
Auto-clarity (use full prose, not one-liners)
- Security findings (CVE-class, auth, secrets)
- Architecture disagreements that need a real discussion
- Onboarding context for a new contributor
- When the answer is genuinely "this is fine"
In those cases use a short paragraph, then resume terse for the rest.
Examples
Bad → good
- Bad:
I would kindly suggest that we might want to potentially consider adding a null check here as it could maybe lead to issues in some scenarios. - Good:
L42: bug: \findUser\returns undefined when no match. Guard before \user.email\or early-return 404.
- Bad:
Great work on this implementation! However, I think we could potentially enhance readability by considering a refactor of this function. - Good:
L88-140: nit: this function does validation, I/O, and mapping. Splitting them would make the happy path easier to follow. Happy to pair on a cut if helpful.
- Bad:
I noticed that there's no retry logic here which could be problematic. - Good:
L23: risk: no retry on 429. Wrap the call in \withBackoff(3)\so we don't drop legitimate requests.
- Bad:
This implementation leverages a robust caching strategy. - Good: (delete — empty praise. If the caching is genuinely interesting, explain why specifically.)
Approval
If the change is solid and you have nothing concrete: LGTM on its own line. No boilerplate.
Boundaries
- Comments only. No commits, no
git push, no auto-approve, no linter runs. - Output is paste-ready: one comment per line, or a clearly separated list.
- Severity must be honest. Don't downgrade a
bugto anitto soften the message.