做什么
检查工作,发现漂移前。 另一种观点认为,早期捕捉到错配. 在自然暂停点、公关前、或感觉不对的时候使用.
技能库 智客分类:Agent 工作流 review
检查工作,发现漂移前。 另一种观点认为,早期捕捉到错配. 在自然暂停点、公关前、或感觉不对的时候使用.
官方网址:skills.sh
先看中文介绍;官方 description 原文单独保留,不改写 SKILL.md。
检查工作,发现漂移前。 另一种观点认为,早期捕捉到错配. 在自然暂停点、公关前、或感觉不对的时候使用.
在以下情况下启动`/审查 ' :在自然暂停点前,感觉有些不对劲,范围似乎已经扩大,或在公关之前
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、When to Use、Review Lenses、The Review Process、Step 1: State the Original Aim、Step 2: Check Alignment。 其中含规范建议的小节:分步指令。
文件分析:除 SKILL.md 外,正文引用了 references/code.md、references/writing.md、references/learning.md,属于带资源的技能包,这些文件按需再读。
Check work and detect drift before committing. A second opinion that catches misalignment early. Use at natural pause points, before PRs, or when something feels off.
/reviewWhen to UseReview LensesThe Review ProcessStep 1: State the Original AimStep 2: Check AlignmentStep 3: Check the FrameStep 4: Detect DriftStep 5: Decide Next ActionOutput FormatReviewAlignment Check
来源分类:skills.sh agent-skill
namereviewdescriptionreferences/code.mdreferences/writing.mdreferences/learning.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 ↗
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。
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-dd14356282303626-%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/open-horizon-labs/skills 取完整目录,不要只建一个 SKILL.md。
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/open-horizon-labs/skills' --list
npx skills add 'https://github.com/open-horizon-labs/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。
A second opinion before committing. Checks alignment between intent and execution, detects drift, and decides: continue, adjust, or salvage.
Part of the Intent → Execution → Review loop at every scale — from a single commit to a multi-week project. Review closes the loop.
Invoke /review when: before committing, at natural pause points, something feels off, scope seems to have grown, or before PRs.
Do not use when: You're in deep flow and making progress. Review at natural breaks, not arbitrary intervals.
Default to the core workflow review in this file. When the artifact being reviewed needs a sharper domain lens, load the matching supporting reference:
Use these as lenses, not replacement workflows. The final review still answers the core /review questions: aim, alignment, sufficiency, mechanism, drift, completion, and next action.
"The aim was: [original intent in one sentence]"
If you can't state the aim clearly, that's the first finding.
Ask six questions. Keep them artifact-neutral; load a review lens only when the artifact needs domain-specific judgment.
If RNA MCP is available, check work against guardrails (oh_search_context with artifact_types: ["guardrail"]) and declared outcomes (outcome_progress).
Before naming ordinary drift, ask whether the interpretation itself collapsed:
Frame Collapse — the working interpretation no longer explains the evidence, constraints, or aim. Treat this separately from drift. Route it by where the break lives: back to /problem-statement for a bad frame, /problem-space for a bad situation map, /aim for changed intent, or /salvage when the session needs learning extraction before restart.
A named alternate frame that remains model-checkable but untested is incomplete execution, not automatic frame collapse and not a note for later. A named risk with a non-adversarial check is also incomplete: if the evidence would pass the tempting local patch, the risk was not retired. Route to Adjust when the selected approach is still plausible and the missing/adversarial check can be added. Route to Reframe only when evidence already invalidates the selected approach or the alternate frame better explains the problem than the current one.
Drift is the gap between where you started and where you are. Explicitly name any drift found:
Scope Drift — task grew beyond original boundaries Solution Drift — approach changed from plan Goal Drift — aim itself shifted without explicit decision Frame Collapse — evidence invalidated the story that made the work make sense
For each drift or frame collapse:
Drift: [type]
Started as: [original frame/plan]
Became: [current evidence/model]
Impact: [what this means]
Route: [continue | adjust | /problem-statement | /problem-space | /aim | /salvage]
| Decision | When | Action |
|----------|------|--------|
| Continue | Aligned, on track, and named risks retired or explicitly accepted | Proceed with confidence |
| Adjust | Minor drift, recoverable issue, model-checkable named risk is still untested, or retirement evidence is non-adversarial while the selected approach remains plausible | Add the missing/adversarial check or correction and continue |
| Reframe | Frame collapse, invalidation triggered, or evidence shows an alternate frame better explains the problem than the selected approach | Return to /problem-statement, /problem-space, or /aim |
| Pause | Unclear aim, major questions, or required human/domain verification | Stop, clarify, then resume |
| Salvage | Significant drift or failed frame needs extraction | Extract learning with /salvage |
If the review surfaces new hard constraints, record them (oh_record_guardrail_candidate).
## Review
**Aim:** [original intent]
**Status:** [Continue / Adjust / Reframe / Pause / Salvage]
### Alignment Check
- Necessary: [Yes/No - brief note]
- Aligned: [Yes/No - brief note]
- Sufficient: [Yes/No - brief note]
- Mechanism clear: [Yes/No - brief note]
- Changes complete: [Yes/No - brief note]
- Risks retired: [Yes/No - every named risk/stop trigger is covered by evidence, accepted with rationale, or triggered]
### Frame Check
[State whether the current frame still explains the evidence. If collapsed, name the new frame or route to reframe/salvage.]
### Drift Detected
[List any drift found, or "None detected"]
### Risk Retirement Review
[List each named risk / alternate frame / stop trigger from Solution Space and Execute with its evidence. Mark untested model-checkable items or non-adversarial checks as Adjust unless observed evidence invalidates the selected approach; mark those as Reframe.]
### Needs Human Verification
Claims that cannot be self-checked by the model and require independent human judgment:
- [claim or assumption that needs external confirmation]
- [one-way decision that was not independently challenged]
- [delivered characteristic that was asserted but not externally observed]
If this section is empty, say why. Model-generated review of model-generated work shares the same blindspots. The purpose of this section is to route the right questions to the actual human checkpoint (PR review, stakeholder sign-off, manual test) rather than pretending the model can answer them.
### Decision
[Reasoning for the status decision]
### Next Steps
[Concrete actions to take]
When the user or agent claims work is complete, verify:
Intent clear? — Can you state what changed and why in one sentence?
Work reviewed? — Has the actual artifact been checked against the stated intent?
Evidence present? — Are the relevant checks, examples, or observations recorded?
Feedback handled? — Have review comments or open objections been resolved or explicitly deferred?
Declared criteria delivered? — Were promised characteristics verified, not merely claimed?
Needs Human Verification surfaced? — Have claims the model cannot self-check been flagged for human attention?
Risk retirement complete? — Were all named risks/alternate frames/stop-pivot triggers tested, explicitly accepted, or routed to Adjust/Reframe/Pause? If incomplete:
"Completion gate: [missing step]. Run the check before marking complete."
If a session file is in use, replace or append ## Review. Preserve frame check, verification result, risk-retirement review, human-verification needs, and unresolved objections for later phases.
Agent 工作流
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
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
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
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…