AgentSkillsCN

critique

为计划、规格或任务成果提供统一的评审意见。自动识别目标对象,派出评审人员,并在任务“就绪”时将其提交至数据库。适用“评审”“审查”“这是否已经就绪?”等指令。

SKILL.md
--- frontmatter
name: critique
description: Unified critique for plans, specs, OR task artifacts. Auto-detects target.
  Spawns auditors. Commits tasks to DB when READY. Use "critique", "review", "is this
  ready".
argument-hint: <project-name|artifact-dir> [--skip-skeptic] [--force-necessity]
inherit:
- review
tools:
- auggie
- warpgrep
spawns_subagents: true
ultrathink: true
required_todos:
- detect-target
- read-target
- check-regression-round-2-only
- spawn-auditors
- resolve-auditor-conflicts
- synthesize
- publish
<role> WHO: Adversarial reviewer ATTITUDE: Every plan solves an imaginary problem until proven otherwise. </role> <purpose> Detect target type (plan/specs/task), spawn auditors who explore CODEBASE (not just read text), verify claims, synthesize verdict. For tasks: commit to DB when READY. </purpose> <workflow>

Steps declare dependencies via consumes/produces frontmatter. Execute steps whose inputs are all satisfied — parallel when independent.

Setup (sequential)

StepConsumesProduces
detect-targetuser-requesttarget-type, target-paths
read-targettarget-type, target-pathstarget-content, claims

TASK PATH (if target_type == TASK)

→ Read and follow: ~/.claude/skills/critique/steps/task-critique.md STOP — task path complete.

PLAN/SPECS PATH

StepConsumesProducesNotes
regression-checktarget-contentregression-findingsRound 2+ only
spawn-auditorstarget-content, target-typeauditor-findingsParallel auditors
resolve-conflictsauditor-findingsverified-findings
synthesisverified-findings, target-contentcritique-verdict

→ Execute in dependency order. For each step:

  1. Read ~/.claude/skills/critique/steps/<name>.md
  2. Complete it fully before reading the next step
</workflow>

Protocols

!cat ~/.claude/skills/_shared/review.md

<rules> - Auto-detect: task.md in one-offs → TASK, specs exist → SPECS, else → PLAN - TASK path: verify paths + Context symbols, commit to DB if READY - **Context = constraint:** If task lists "MUST use: X at file:line", verify X exists via warpgrep. Worker must use it. - PLAN/SPECS path: full auditor workflow - Auditors MUST explore codebase, not just read text - Auditors spawn in SINGLE message (parallel) - Same fix pattern = 1 blocker with count - Verify claims BEFORE synthesis — auditors lie about tool behavior - De-dupe overlapping findings — merge, don't stack - **Verdict routing:** Ask user to confirm next action (APPROVED → /decompose or ft epic decompose | FIX_AND_SHIP/REVISE → /revise) </rules>