Skill v1.0.1
currentAutomated scan100/100+6 new
version: "1.0.1" name: cross-review description: Give an existing proposal an independent second opinion when requested; return one judged recommendation. disable-model-invocation: true
Cross Review
Challenge an existing proposal when the user requests a second opinion. Return one keep, simplify, replace, or needs-decision verdict. The output is that verdict: the proposal, implementation, and any new plan stay with their owners. Use the appropriate code-review skill for PRs.
Freeze The Input
Identify the exact proposal, the problem it claims to solve, and any explicit constraints or non-goals. Prefer the latest assistant proposal when the user says "this" or "that proposal." Ask one concise question only when multiple candidate proposals make the target materially ambiguous.
Freeze that input for the round: every reviewer sees the same proposal and none sees another's findings, which is what keeps the passes independent.
Read reviewers when choosing and dispatching independent perspectives. Preserve the scope of each selected skill.
Judge Findings
The main session classifies each material finding:
accept: changes the recommendation;merge: duplicates or strengthens another accepted finding;reject: weak evidence, wrong scope, taste, or reviewer-boundary leakage;park: plausible but outside the stated problem;needs-decision: changes product intent, a public contract, safety, privacy,
cost, dependency policy, or another user-owned boundary.
Choose the verdict:
keep: no material finding justifies changing the proposal;simplify: preserve the approach but remove or narrow real complexity;replace: another approach better satisfies the same goal and constraints;needs-decision: no honest recommendation exists without user direction.
Judge findings on evidence rather than averaging reviewers, and keep an alternative only when it wins on the original goal. Prefer the smallest recommendation that satisfies the goal and proof needs.
Converge
Stop after one round unless accepted findings materially revise the proposal and another pass could change the verdict. In round two, send the revised proposal only to reviewers whose concerns were affected. Two rounds is the limit; a request for a broader planning exercise goes to $agent-planning-loop.
Stop when the proposal is kept, the revision resolves material findings, the same verdict repeats, remaining items are implementation details, or a needs-decision item blocks judgment.
Output
Return a compact decision packet rather than reviewer transcripts:
Cross-review verdict: <keep | simplify | replace | needs-decision>Proposal reviewed: <one sentence>Accepted findings: <material findings>Recommended proposal: <revised proposal, or original when kept>Recommended next action: <stop | preflight | planning route | named review route>
Add rejected/parked items, a material disagreement, or a note that review ran inline rather than with independent workers only when they apply.
If the proposal is already lean, say so and stop; keep is a complete result.