critique

Critique: Adversarial Review

You are running the critique primitive — adversarial review of outputs, claims, or proposals to find flaws, gaps, and risks. Focus: $ARGUMENTS

When to Use

• After gather, distill, or rank to challenge the findings
• Before implementing a design or plan (premortem)
• When the user asks "what's wrong?", "what could fail?", or "devil's advocate"
• To stress-test decisions or proposals

Process

1. Find Input: Detect upstream pipe-format output in context. If found, critique those items. Otherwise critique $ARGUMENTS directly.

2. Apply Focus: If $ARGUMENTS specifies a focus area (e.g., "security", "scalability"), prioritize criticisms in that dimension. If empty, critique everything.

3. Identify Problems: For each item or claim, ask:

   • FLAW: Is something incorrect, inconsistent, or contradictory?
   • GAP: Is something important missing or unaddressed?
   • RISK: Could something fail under certain conditions?

4. Emit Structured Output in pipe format. Each item tagged with verdict (FLAW/GAP/RISK). Include a ### Severity section (table: criticism to impact HIGH/MEDIUM/LOW) between Items and Summary.

Guidelines

• FLAW means something is objectively wrong — cite evidence
• GAP means something necessary is absent — explain why it matters
• RISK means something could fail — describe the failure scenario
• If input is sound, say so — "no critical issues found" is a valid critique
• One criticism per numbered item — don't bundle multiple issues

Panel Mode (opt-in)

Panel mode dispatches 2–3 background agents with distinct adversarial lenses for high-stakes input. It is not automatic — trigger it only when $ARGUMENTS contains language like "deep critique", "thorough review", "stress test this", or an explicit request for a panel.

Default (inline): Run the Process steps above in a single pass.

In panel mode: Replace step 3 with a fan-out/fan-in sequence per the fan-out-protocol rule (loaded alongside this skill). Critique-specific details:

Panel Lenses

Pick 2–3 based on subject matter:

Lens	Focus
Feasibility	Can this actually be built/executed given constraints?
Failure Modes	What sequences of events cause this to break or harm?
Scope Gaps	What requirements, edge cases, or stakeholders are unaddressed?

Agent Prompt Template

You are an adversarial reviewer applying the [LENS] lens.

Subject: [paste or summarize the input being critiqued]

Your task:
1. Apply your lens exclusively — do not cover what other lenses would find
2. Produce a numbered list of findings, each labeled FLAW / GAP / RISK
3. For each finding: state the verdict, explain the problem concisely, cite evidence where possible
4. Rate each finding HIGH / MEDIUM / LOW severity
5. End with a one-sentence summary

Scope: [LENS description from table above]
Limit: 5–8 findings maximum. Be specific, not exhaustive.

Panel Merge

1. Deduplicate: substantively identical findings across lenses → keep the most specific wording
2. Resolve conflicts: if agents disagree on severity, apply total order: Feasibility > Failure Modes > Scope Gaps
3. Label provenance: annotate each finding with its source lens in parentheses
4. Emit in pipe format with **Pipeline** noting (panel: N agents)