dancing-bear: Skill for Claude Code

Critique — Adversarial Review Entry Point

Runs workflows/shared/critique.yaml, which wraps the existing workflows/shared/plan-critic.yaml fragment (validate → critic (Opus) → validate-critique → consolidate → human-gate → revise) behind a prepare-target stage so it can be pointed at a single doc, several files, or the current diff — not only used as an include: inside another workflow.

When to Use

• User says "critique this", "adversarial review", "opus critique", "run a critic pass"
• Before running a normalize/apply workflow that writes code across many files
• Reviewing a design doc or workflow YAML before committing to it
• Reviewing a batch of just-written code/docs before opening a PR

Usage

/critique .llm/DESIGN_CRITERIA.md
/critique workflows/code/design-criteria-normalize.yaml
/critique --targets .llm/DESIGN_CRITERIA.md,workflows/code/design-criteria-audit.yaml,workflows/code/design-criteria-normalize.yaml
/critique --diff                        # uncommitted changes (git diff HEAD)
/critique --diff-ref main...HEAD        # diff against a ref
/critique --focus "worktree isolation safety"

Input Parsing

Argument	Required	Default	Description
positional / --target	No*	--	Single file path to critique
--targets	No*	--	Comma-separated file paths, concatenated into one critique target
--diff	No*	--	Critique uncommitted changes (git diff HEAD)
--diff-ref	No*	--	Critique a diff against a ref, e.g. main...HEAD
--focus	No	full-spectrum	One-line directive narrowing the critique (e.g. "idempotency on re-run")

* Exactly one of --target/positional, --targets, --diff, --diff-ref should be given. If none are given, defaults to --diff (uncommitted changes) — this matches "critique what I just did."

Running

Invoke via the /workflow skill — do not hand-roll a critic Agent() call when this skill applies, since the included fragment already chains fact-checking and a human gate around the raw critique (a lone critic-agent spawn skips both). Do NOT invoke ./bin/workflow run workflows/shared/critique.yaml directly: agent-required stages route through SkillDispatcher, which only writes dispatch JSON and returns pending — it does not spawn or wait for agents, so the critique will not actually execute end-to-end. The /workflow skill is the orchestrator that spawns agents, waits for results, and drives human gates.