菜单
Review the latest work unit with fresh, unnarrowed eyes and surface any problem it contains as follow-ups. Use after a work-unit commit, or when the user runs /agent-review.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Continue working on a story or task by reading progress.md, executing the next work unit, and updating progress. Use when user wants to resume story/task work or mentions /continue command.
Run a Selenium test in visible (non-headless) mode with slowdown so the user can watch it. Use when user wants to demo or visually watch a Selenium test or mentions /demo command.
Preview the planned design for a scenario before writing tests. Shows domain model changes, usecase patterns, and key design choices. Use before red-usecase to get user approval on the implementation approach. If the user rejects the design, offers to discuss inline or escalate to /architecture for a full ADR.
Write a journey summary capturing noteworthy moments (predictions, decisions, surprises, mistakes, quirks). Runs automatically the moment the agent observes something worth noting during work, and manually before /clear or /compact as a final sweep. Idempotent — skips entries already recorded. Use when the agent observes a noteworthy moment, before clearing context, or when the user mentions /handoff.
Re-explain the last thing in plain words. Use ONLY when the user explicitly types /plain or asks to "explain that simply / in plain words / like I'm not an expert". Never invoke on your own initiative — this is a manual button the user presses when something was hard to follow.
Imagine the incident a just-shipped work unit would cause and surface unguarded gaps as follow-ups. Use after a work-unit commit, or when the user runs /premortem.
| name | agent-review |
| description | Review the latest work unit with fresh, unnarrowed eyes and surface any problem it contains as follow-ups. Use after a work-unit commit, or when the user runs /agent-review. |
A fresh-context pass that reads the latest work cold and surfaces any problem the
work contains. Distinct from /premortem: that imagines what the work is
missing; this audits what it holds. The two are complementary lenses, not a
repeated read.
/agent-review — review the latest work-unit commit (HEAD)/agent-review <ref> — review a specific commit or range.claude/agents/agent-review-agent.md.HEAD); use <ref> if given.agent-review-agent (Agent tool) with the diff and one line of
context on what the work unit did.When an orchestrator (/continue) dispatches this as a work-unit review pass, the
agent runs concurrently with premortem — two independent fresh-context reads
of the same diff. The timing and the diff it reads in that path (the pre-refactor
behavior commit — HEAD or HEAD~N..HEAD — dispatched in the /refactor batch)
are owned by /continue; see its
"Pre-Commit Review Passes" section. Run standalone, this skill reviews HEAD (or
<ref>) on demand. The agent is read-only; it reports, it does not edit.