agentsclimarketplace

Teamwork review

Skill JinPLu/Teamwork/plugins/teamwork-skill/skills/teamwork-review

Evidence-first skills for Codex, Cursor, and Claude Code: research, debugging, planning, execution, review, and goal loops.

Install
npx -y skills add JinPLu/Teamwork --skill teamwork-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 8 stars8 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when the user asks to review, audit, critique, validate, or decide whether a candidate or claim is correct or complete, or a named material risk gate requires independent review of a sealed integrated candidate; do not use for each Worker slice, to implement fixes, write a plan, or perform ordinary evidence collection.

SKILL.md

4.9 KB, as published. Nobody here has run it

Teamwork Review

Issue an evidence-based ACCEPT, REVISE, or BLOCKED verdict. Review is read-only: do not edit the candidate, apply fixes, publish, or perform external effects even when a fix seems obvious.

Each Worker self-verifies its owned slice. Do not review each Worker slice or code delta independently. Root integrates authorized changes and seals one stable candidate with scope and direct evidence. Run one independent initial pass on that sealed integrated candidate only for user-requested review or a named material risk gate. A named risk gate may review its exact protected boundary before integration when delay would invalidate proof.

Method

  1. Establish candidate, scope, acceptance criteria, protected boundaries, and evidence needed for the verdict. User-requested review takes precedence over inferred criteria; do not invent requirements.
  2. Inspect primary evidence directly: source and diff, tests and configuration, runtime output, rendered artifacts, or authoritative external sources as the candidate requires. A summary or claimed test result is input, not proof.
  3. Check correctness first: acceptance criteria, security/permission boundaries, data behavior, regressions, error paths, compatibility, and direct real-path evidence. Do not let style or cleanup substitute for this pass.
  4. Then inspect only the changed scope for cohesion and deslop: wrong-layer or duplicate owners, thin wrappers, dead code, speculative or single-consumer abstractions, unnecessary compatibility or parallel modes, broad catches, masking fallbacks, unnecessary public surface, temporary residue, and stale touched comments or configuration. A justified multi-file boundary, an explicitly accepted product fallback, and purely pre-existing debt are negative controls, not blockers.
  5. Give each finding a stable R-* ID and classify it once:
    • BLOCKER: a failed acceptance criterion, regression, boundary breach, or missing evidence required for the verdict;
    • FOLLOW-UP: a real non-blocking issue outside the accepted result;
    • SUGGESTION: an optional improvement.
  6. State the concrete evidence, affected criterion or user impact, and smallest correction route. Do not promote an out-of-scope improvement into a blocker unless the current acceptance criteria cannot pass without it.

Load references/strict-review.md only for a named strict release, security, permission, data, destructive-risk, or public-contract gate.

Failed, blocked, partial, and unverified findings change only with new direct evidence. Combine initial findings into one repair batch. The same Reviewer may perform at most one bounded delta recheck per candidate, limited to stable findings and fix-introduced regressions. Any source change after that creates a new candidate; expanded scope requires a fresh review decision. Root retains final acceptance.

Reviewer always stays read-only. In an initialized writable project, every verdict defaults to a review/conclusion artifact unless the user says no files, off-record, read-only, no writes, or equivalent. Freeze the terminal verdict before persistence. Root freezes the verdict and Reviewer returns a bounded packet: purpose/audience, facts/sources, frozen decision/status, style/structure, artifact kind/consumer, preserve/forbid, findings, evidence, verdict, and residual risk. Writer uses artifact-inspect -> artifact-schema <create|update|supersede> -> artifact-apply; the transaction derives the destination and registers the ordinary index. Writer is disposable; Root may continue only answer-invariant delivery work and must join before claiming saved or durable. Interruption before artifact-apply means unsaved unless surviving evidence permits a new frozen packet. Missing project memory, Writer, brief, authority, consumer, or transaction blocks only persistence: deliver the verdict and report it unsaved/blocked. No Reviewer, Root, or Worker fallback writes it. Persistence does not imply Root/user acceptance.

Lead with blockers ordered by severity and include precise file/line or artifact locations when available. If there are no findings, say so explicitly. ACCEPT requires direct support for every material criterion and no open blocker; REVISE means correctable blockers remain; BLOCKED means required evidence or access is unavailable. Reviewer never asks the user; propose missing user-providable evidence to Root before declaring BLOCKED, but make no change. Candidate changes invalidate the old review and require a new sealed candidate.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.