跳到主内容
智客 ZICQ

技能库 智客分类:Agent 工作流 review

Review

审查当前的分支和未承诺的更改:先读出实际的diff, 检查更改的测试和CI, 报告风险和有结构反馈的证据(Blockers / Conservation / Nits),然后在承诺前应用经批准的修改并询问.

75 安装量

官方网址:skills.sh

技能介绍

先看中文介绍;官方 description 原文单独保留,不改写 SKILL.md。

做什么

审查当前的分支和未承诺的更改:先读出实际的diff, 检查更改的测试和CI, 报告风险和有结构反馈的证据(Blockers / Conservation / Nits),然后在承诺前应用经批准的修改并询问.

何时用

官方 description 未单独写出 Use when。按规范,代理会在用户任务与这段 description 的关键词匹配时激活本技能。

代理如何加载

按 Agent Skills 渐进披露:启动时只加载 name 与 description(约 100 token);任务匹配后才读入整份 SKILL.md 正文;scripts/、references/、assets/ 仅在需要时再读。 本文件正文结构:/review、Workflow、Scope Discipline、Security-Before-Push、Agent-Authored Diffs、Severity Tiers。 其中含规范建议的小节:分步指令。

文件分析

文件分析:除 SKILL.md 外,正文引用了 references/agentic-code-review.md、references/code-review-and-quality.md,属于带资源的技能包,这些文件按需再读。

官方 description(原文)

Reviews current branch and uncommitted changes: reads the actual diff, checks changed tests and CI first, reports risk and evidence with structured feedback (Blockers / Concerns / Nits), then applies approved fixes and asks before committing.

/reviewWorkflowScope DisciplineSecurity-Before-PushAgent-Authored DiffsSeverity TiersOutput FormatReview ContextBlockersConcernsNitsQuestions

来源分类:skills.sh agent-skill

SKILL.md 与 Agent 调用

官方规范 ↗
name
review
description
Reviews current branch and uncommitted changes: reads the actual diff, checks changed tests and CI first, reports risk and evidence with structured feedback (Blockers / Concerns / Nits), then applies approved fixes and asks before committing.
  1. 发现技能客户端向 Agent 提供名称与描述目录。
  2. 匹配与调用用户指定或任务匹配后,载入 SKILL.md 指令。
  3. 按需加载按步骤读取参考文档、使用脚本与素材。
指令中引用的文件 · 2
  • references/agentic-code-review.md
  • references/code-review-and-quality.md

以下路径提取自原文;文件是否齐全请以来源仓库中的完整目录为准。

具体调用语法与可用工具以目标 Agent 客户端为准。 查看调用机制说明 ↗

安装这个技能

Skills CLI ↗

先选择目标 Agent 和安装范围,保留技能包的附属文件,安装后检查客户端能否发现该技能。

该技能引用了附属文件,请从来源获取完整目录;仅复制 SKILL.md 可能缺少依赖。

交给 Agent 安装

复制安装指令给支持 Agent Skills 的代理,确认其中的目标目录与客户端匹配。

把 Agent Skill「review」安装到我的项目:SKILL.md 原文与官方 description 见 https://zicq.com/zh/skills/skl-d0362313028b4ef7-%E5%AE%A1%E6%9F%A5.html
请存为 .cursor/skills/review/SKILL.md 或 .claude/skills/review/SKILL.md,frontmatter 的 name 与 description 保持原样,不要改写。
该技能还带 scripts/、references/、assets/ 等文件,请从 https://github.com/bsene/skills 取完整目录,不要只建一个 SKILL.md。

GitHub 完整包 ↗

终端安装 · Skills CLI

需要 Node.js 与 npx。先查看仓库技能列表,确认实际名称。

npx skills add 'https://github.com/bsene/skills' --list

npx skills add 'https://github.com/bsene/skills' --skill 'review'

CLI 会交互选择目标 Agent,默认安装到项目;用户级安装使用 -g。先通过查看命令核对仓库内容,再用 npx skills list 检查已安装技能。

阅读排版
--- name: review description: > Reviews current branch and uncommitted changes: reads the actual diff, checks changed tests and CI first, reports risk and evidence with structured feedback (Blockers / Concerns / Nits), then applies approved fixes and asks before committing. TRIGGER when: user says "/review", "review this PR", "review my changes", "code review", "review the diff", "PR review", "review branch", "what do you think of my code", "look at my changes", "give me feedback on this". DO NOT USE when: user wants to understand what changed without requesting feedback; user asks for explanation only; there is no supplied or local diff to review. --- # /review ## Workflow 1. Get the actual scope, never reconstruct it: review a supplied diff; otherwise inspect `git diff main...HEAD`, `git diff --cached`, `git diff`, and `git status --short`. This includes tracked uncommitted work by default. If untracked files exist, ask the user whether to include them; state which parts were available. 2. Read changed tests and CI/config before implementation. A behavior-changing test rewrite without a sound reason, or an unjustified weaker gate, is a Blocker. 3. Produce structured review with risk, blast radius, and available test/validation evidence. When changed code has obvious callers, dependents, or public surface, recommend checking them and their relevant tests; do not turn every review into call-graph archaeology. 4. Ask ≤5 grill-me questions — one per key design decision; design intent only, not style. 5. On approval of specific fixes: apply them → run relevant tests and lint → stop and grill the user about remaining design or validation questions → ask explicit approval before committing. ## Scope Discipline Review **only what changed** — never report pre-existing, unmodified code as a finding. The diff is the contract, but it is not the only context: read a changed file or an obvious caller, dependent, or relevant test when needed to establish the changed code's impact. Real repos make this a hard rule: > Focus findings exclusively on changed code. Use surrounding code only to understand impact; > do not turn a review into a backlog of pre-existing issues. ## Security-Before-Push For anything about to be pushed, run a **security-focused review of the branch's own changes** before shipping (optional `trivy fs` on vulnerable deps). Watch specifically for: logged secrets/tokens/PINs/PII, and `__proto__`/prototype-walking reads on untrusted config. ## Agent-Authored Diffs Run the test-tampering and CI-weakening checks in every review. If the branch was authored largely by an agent — or you suspect it was — also layer [agentic-code-review](references/agentic-code-review.md). For every new or changed AI/LLM feature, inspect untrusted prompt input: it is a Blocker when it can influence privileged data or actions, otherwise a Concern. Tier review depth by blast radius, not by author; the human who clicks merge owns the change. ## Severity Tiers Classify the observed behavior, not a label: an incorrect calculation is a Blocker; a name or rename is a Nit unless the supplied diff or context proves it breaks compatibility or behavior. | Tier | Criterion | Required Action | | ----------- | ----------------------------------------------- | -------------------- | | **Blocker** | Incorrect behavior, security hole, data loss, broken invariant, or unjustified weakened gate | Fix before merge | | **Concern** | Design flaw, missing evidence/test, or perf issue without demonstrated user harm | Discuss + likely fix | | **Nit** | Style, naming, minor duplication | Optional | If a performance finding may affect availability but its impact is unclear, ask the user whether to treat it as a Blocker before assigning a tier. ## Output Format ``` ## Review Context - **Scope:** branch diff + staged + unstaged changes - **Risk:** Medium - **Blast radius:** `createInvoice()` feeds the payment record - **Validation evidence:** no test output supplied ## Blockers - `src/auth.ts:42` — expiry check uses `<` not `<=`; off-by-one passes expired tokens ## Concerns - No test for error path in `processPayment()` ## Nits - `userData` → `user` (same meaning, shorter) ## Questions 1. Why is `retryCount` hardcoded to 3 — should this be config? ``` ## Deep-Rigor Mode When the user asks for a strict / thorough review (e.g. "/review --strict", "deep review", "rigorous review", "code quality audit"), layer the [code-review-and-quality](references/code-review-and-quality.md) rubric on top of the workflow above: walk the five axes (correctness, readability, architecture, security, performance), promote its presumptive blockers into `Blockers`, and hold commits to its higher approval bar. For agent-authored branches, also layer [agentic-code-review](references/agentic-code-review.md). ## Related - [mattpocock/code-review](https://skills.sh/mattpocock/skills/code-review) — two-axis review (Standards + Spec vs the originating issue) run as parallel sub-agents. Source: - [code-review-and-quality](https://github.com/addyosmani/agent-skills/tree/main/skills/code-review-and-quality) — upstream of the vendored deep-rigor rubric used in Deep-Rigor Mode above; see also [agentic code review](https://addyosmani.com/blog/agentic-code-review/) for the agent-diff failure modes. ## Benchmark Scenario: `.benchmarks/scenarios/review-001-severity-tiers.md` · Run: 2026-08-31 · Log: `.benchmarks/runs/2026-08-31/review-001-severity-tiers.json` | Model | Without | With | Delta | | ----------------- | ------- | ---- | ----- | | claude-opus-4-8 | 67% | 83% | +16% | | claude-sonnet-4-6 | 83% | 67% | −16% | | claude-haiku-4-5 | 83% | 83% | +0% | > **NEG (run 2026-08-31)**. Sonnet −16 on the severity-tier boundary; opus +16. Classification ambiguous between criteria and tier table — diagnostic c5-only re-run queued before any edit (cap reached). Gate per `.agents/skills/skill-optimizer/rules/release-gates.md`.

相关技能

Agent 工作流

Skill Creator

创造有效技能指南。 当用户想创造出新的技能(或更新现有的技能),以专业知识,工作流程,或工具集成来扩展克洛德的能力时,应该使用这种技能.

Agent 工作流

Clawdhub

使用ClawdHub CLI搜索,安装,更新并发布从taladhub.com的代理技能. 需要获取苍蝇上的新技能时使用,将安装的技能同步到最新版本或特定版本,或者发布 npm-instainddhub CLI 的新/更新的技能文件夹.

Agent 工作流

Agent Team Orchestration

管弦乐团多代理团队,任务设定周期,交接协议,审查工作流程. 使用时间: (1)建立2+特派员队伍,具有不同专业,(2)确定任务路线和生命周期(收录框_ spec_建设_审查_完成),(3)在特派员之间制定交接协议,(4)建立审查和质量关口,(5)管理特派员之间的交流和文物共享.

Agent 工作流

Superpowers

Spec-first,TDD,子代理驱动的软件开发工作流程. 当:(1)构建任何新功能或应用——触发脑暴_计划_子代理执行回路,(2)调试出一个bug或测试失败——触发系统性的根起过程,(3)用户说"让我们构建","帮助我计划","我想添加X",或"这个被打破",(4)完成一个功…