Immutable. This exact content is served forever at /api/v1/blob/db73ea3d1b84a45b.
--- name: toxic-review description: Use for a strict, skeptical, risk-focused review of a named work product without attacking people or inventing criticism. --- # Hope Toxic Review Be demanding about the work and respectful toward people. Use the active host session to bind one named work product, assign fresh reviewers, validate consequential or uncertain findings when needed, adjudicate the findings, and report the result. Do not let the active session act as a reviewer or finding verifier. ## Bind the target State the target, current stage, material risks, evidence in scope, and evidence that is unavailable. Do not widen the review into unrelated work. Changed evidence starts a new review. ## Choose red reviewer roles Choose the smallest useful role set: use one focused role when it covers the material question, and add roles only to test distinct risks. Use a fresh subagent context for every red reviewer role, including a one-role review. Give it only the exact target, its role and risks to test, direct evidence it may use, explicit exclusions, expected output, and the location of this Skill. Tell each reviewer to read this Skill before acting. Do not let a reviewer inherit the active conversation, previous reasoning, drafts, implementation narrative, prior conclusions, or another red reviewer's input or output. Run roles in parallel or in separate sequential contexts. Repeated prompts in one shared context are not independent reviewers. If a fresh context is unavailable, stop without performing the review. ## Review the work Each finding needs: - a concrete issue; - practical impact; - evidence; - a proposed action; - priority; - confidence; and - an important limit or uncertainty. Use the target's priority vocabulary when it exists; otherwise use high, medium, or low. Do not assign a release-blocking label unless the available evidence shows that the work should stop. Do not manufacture criticism. Finding no material issue is a valid result. Do not turn uncertainty into an established defect. Use the causal-completeness method only when the named work product makes or relies on a material causal claim. When it applies, read [references/causal-review.md](references/causal-review.md) and assign the method to one role. Do not activate it merely because the target is an incident. ## Validate findings when needed Seal the red reviewers' candidate findings before validation. Use a blue verifier subagent in a fresh context when a candidate finding: - would stop a release or require costly or destructive action; - relies on incomplete or ambiguous evidence; or - has materially uncertain scope or impact. Review size alone does not call for a blue verifier. A small review may remain red-only only when no candidate finding meets these conditions. Give the verifier only the named target, scoped evidence, sealed candidate findings, explicit exclusions, the location of this Skill, and the expected output. Do not give it hidden reasoning, the active conversation, prior conclusions, or unsealed reviewer output. Tell the verifier to read this Skill before acting. The verifier tests the candidate findings. It does not defend the work product or try to cancel criticism. For each finding, it must: - try to falsify the claimed issue, impact, scope, and proposed action; - identify missing context, unsupported assumptions, and overstatement; - acknowledge evidence that survives the challenge; and - return uphold, qualify, or refute with direct evidence and any important uncertainty. Use one verifier for related findings. Add another only when a distinct kind of expertise is needed. If a required fresh verifier is unavailable, stop and state which validation could not be performed. ## Adjudicate Use the active host to combine red reviewer output and any blue verifier output. Adjudicate only from the named target, scoped evidence, red reviewer output, and any blue verifier output. Do not introduce hidden conversation context as evidence. Treat blue verification as evidence for adjudication, not as a decision. Judge each material finding by the target evidence, impact, current scope, feasibility, duplication, and any blue challenge. Resolve red and blue disagreement from the evidence rather than choosing a side. Do not count reviewer votes. Accept, partly accept, reject, defer, or merge each finding. A deferred finding needs a concrete next step. Keep rejected and duplicate findings out of the actionable list. ## Respond Use one strict, competent voice. Lead with the highest-priority accepted issue. Keep deferred risk visible. If no material issue was found, say so and name the checked scope and limits. Do not attack a person. Perform one review round for one evidence snapshot. Start another only when changed evidence or an accepted high-impact finding creates a different material question. Do not create a custom model adapter, private role-run state, evaluation record, or persisted review JSON.