Critique

Run an independent adversarial review of an artifact, evidence, a theme, a problem one-liner, a score, a brief, or a FAQ. The entire value is independence: critique never grades work it produced, and never argues to agreement. Run it as a fresh pass (a subagent with its own context; a different model where available).

Independence (mandatory, read first)

Run as a separate pass with its own context, ideally a different model, so the agent that produced a finding never grades it. A critic sharing the author's context tends to rubber-stamp, which defeats the purpose. Treat the author's reasoning as something to attack, not a position to defend. State in the output that it ran independently.

Mechanism: spawn each critic/lens pass as a Task-tool subagent (a general-purpose agent) whose context contains only the artifact under review plus the relevant rubric, not the author's reasoning chain. For high-stakes reviews, prefer a different model. The bounded panel rounds are driven by the parent: collect each subagent's findings, let the author revise, then re-spawn fresh subagents for the next round.

Three modes, pick by what's under review

1. Transition critic (default, lightweight)

Run at any stage handoff (evidence → themes → problem → score → plan). Apply the relevant checks in ../pressure-test/references/critic-checklist.md plus the always-on cross-stage consistency checks. Return blocking issues (with fixes), suggestions, strengths, and overall readiness (ready / needs fixes / not ready). Every flag must be actionable.

2. Crispness grader (for a problem one-liner)

Score the problem statement on the dimensions in references/crispness-rubric.md (0-10 each), total it, and block on any dimension scoring 0. Return the lowest-scoring dimensions and a one-line rewrite that lifts them. Scale the bar to the stage, an exploratory founder one-liner needs altitude and a named persona, not yet a measured consequence. Flag, don't tyrannize.

3. Multi-lens panel (heavy; for a brief, a FAQ, or a high-stakes finding)

Convene the lenses in references/lenses.md, skeptical customer, competitor, economics/CFO, feasibility/engineer, contrarian, security/legal (only where relevant). Each lens attacks the answers or findings it owns. Run bounded rounds (max 3); stop early when a round raises no new substantive issue. Terminate every point into one of three states, never "consensus":

• Resolved: answered and backed by evidence that clears the gate.
• Open: needs data the panel lacks → emit a specific gap question (with where to look) routed to gather-evidence. Do not argue it closed.
• Contested: a genuine unresolved disagreement → log it as a recorded risk for the human to weigh.

Calibration

The critic flags; it does not block everything. Match severity to stage: early discovery tolerates thin evidence (flag as a gap); a commit decision does not. An over-eager critic that demands strength-5 proof from a day-one idea is miscalibrated.

Relationship to pressure-test

pressure-test is the entry point for attacking a whole existing artifact (a plan/roadmap/PRD/bet), it runs this engine. critique is the general form, usable at any transition, for crispness, or as the panel. One adversarial engine; pressure-test is its plan-level door.

Record

Append findings, the crispness score, and any open/contested items to the investigation's decisions-log.md. Critique informs; it does not silently edit the artifact.

Output

Lead with the verdict. Modes 1-2: blocking / suggestions / strengths / readiness (plus the crispness score and rewrite). Mode 3: each point tagged resolved / open / contested, lead with the count in each state, open items carrying their gap question, contested items carrying the dissent. End with Next move: and recommend revising the artifact, gathering the named missing evidence, or pressure-testing it if it is a plan/roadmap/PRD.