critique

# Critique Review plans and designs, verify model findings against code, or close an implemented plan. Select the mode before loading its workflow or dispatching reviewers. ## Choose the work | Request or current artifact | Route | Read when selected | |---|---|---| | Diff, PR, or “review this diff” | `/diff-review` only | The diff-review skill; do not also critique the same diff | | A new execution route before its first paid or irreversible launch | One repo-access reviewer traces the route's code; not `model`, whose critics cannot see code | `~/.claude/rules/plan-review-gate.md` (“Also trigger”) | | Explicit `model`, or a plan, design or stage/dataset retirement with no later state | `model` | [Adversarial review](lenses/adversarial-review.md) | | Explicit `audit-plan`, or audit backlog (≥15 open items) plus remediation plan | `audit-plan` | [Repo audit plan review](lenses/repo-audit-plan-review.md) | | Explicit `verify`, or findings/audit output with no plan | `verify` | [Verification](lenses/verification.md) | | Explicit `close`, or a recent plan with implementation commits since plan start | `close` | [Plan close review](lenses/plan-close-review.md) | Actual diff review always routes to diff-review. For other artifacts, explicit modes take precedence; with no mode, check close, audit-plan, verify, and finally model. **Choose the smaller workflow before reading further:** - A single specific bug: inspect the code directly. Scientific or external factual claims: use `/research`. Already human-verified findings need no standalone verification pass. - Skip the close workflow for trivial plans (<30 lines, one function, obvious correctness), research/analysis that produces no code, and config/data-only changes with no logic changes. Use the checks appropriate to the artifact. - Audit-plan is the repo-inventory workflow: fresh same-session lane evidence can be reused; otherwise its first audit pass requires repo lanes before critics. It is not the default for an ordinary plan. - Model review is convergent. Use `/brainstorm` for divergent ideation. Extra cosigners, formal review, and broader presets are conditional choices in the selected workflow, not automatic additions. ## Shared review contracts - **One pass per layer:** `review_targets.diff_target` belongs to diff-review; `design_target` belongs to critique. Honor `dispatch.json` layer ownership and blockers. A closeout may need both layers, once each. - **Manifest-driven dispatch:** assemble the actual review packet, run `review_gate.py triage`, then pass its matching `--dispatch-manifest` to `model-review.py`. Explicit CLI flags override manifest fields; do not reuse another packet's manifest. Exact commands, provenance, valid `repo|packet` scopes, and budget rules: [Dispatch](references/dispatch.md). - **Ground repo premises:** a plan depending on live callers, schemas, or joins needs repo evidence before packet-only critics. The premise scout supplies this in normal repo reviews; self-contained packets may disable it. Declare target users, current/designed scale, and rate of change. Include only current governance relevant to this review: [Context assembly](references/context-assembly.md). - **Preserve the evidence:** read every axis output and account for every extracted finding. Inspect `coverage.json` for packet provenance, dispatch coverage, and extraction/verification totals. `verified-disposition.md` checks anchors and local corroboration; it is not semantic proof. - **Verify before acting:** model agreement and self-reported confidence do not establish truth. Confirm code behavior or external facts at their sources. Fix all confirmed/corrected findings within the authorized scope; give a reason per deferred item. Preserve genuine judgment disagreements for the user. - **Migration stance:** default to full caller migration and deletion of old paths. Any retained compatibility boundary needs a live consumer, reason, and removal condition. ## Supporting references Load only the reference needed for the selected task: - [Dispatch](references/dispatch.md): axes, transports, subscription constraints, packet manifests, artifact contracts, and failures. - [Context assembly](references/context-assembly.md): gathering patterns, scope, and context biases. - [Extraction](references/extraction.md): manual extraction, complete disposition, and multi-round coverage. - [Prompts](references/prompts.md): custom prompt templates; the script owns default prompts. - [Biases and anti-patterns](references/biases-and-antipatterns.md): interpreting reviewer errors and disagreements. - [Known issues](references/known-issues.md): read when debugging dispatch; append new incidents here or through `~/Projects/skills/hooks/append-skill-memento.sh critique '<one-line issue>'`. - [History and calibration](references/history.md): preserved rationale and superseded routing; consult for provenance, not current instructions. $ARGUMENTS