About project-review
project-review is a project-level acceptance skill that evaluates an approved target and bounded implementation, then returns PASS, FAIL, or BLOCKED. It freezes the acceptance baseline, coordinates read-only reviewers, and leaves authorized repairs to the Producer.
Key features
- 01Freeze an acceptance baseline and profile
- 02Coordinate read-only reviewers
- 03Validate findings and bounded repairs
- 04Issue PASS, FAIL, or BLOCKED verdicts
- 05Record state for review resumption
Use cases
- 01Check implementation before release
- 02Evaluate a skill and its evidence
- 03Resume an unfinished recorded review
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ light-cheng/ skills --skill project-reviewAfter installing, type/project-reviewin Claude Code to confirm it loaded.
Bundled files
References and scripts installed along with SKILL.md.
agents/openai.yamlreferences/acceptance-charter.mdattribution.mdevidence-protocol.mdfinding-schema.mdmigration.mdmission-center-compatibility.mdreview-rubric.mdstopping-rules.mdsubagent-protocol.mdWORKFLOW.mdreferences/profiles/agent-skill.mdgeneric.mdmanuscript.mdsoftware.mdspecification.mdtests/test_agent_skill_profile_contract.pytest_generic_profile_contract.pytest_manuscript_profile_contract.pytest_software_profile_contract.pytest_specification_profile_contract.pyDetails
From the project READMEProject Review
project-review is the project-level final acceptance capability. It is
model-invoked and may also be invoked by hand. It freezes the acceptance
baseline, drives read-only reviewers through review-loop, and issues the
final PASS, FAIL or BLOCKED. The Producer makes authorized repairs;
project-review never repairs artifacts itself.
When to use
- The project has an approved target (SPEC, brief or ticket graph) and a
bounded implementation to check: after
light-implementor a batch of tickets, beforerelease-workflow. - Skill admission, when the acceptance source is a package's
SKILL.mdand its evidence.
When docs/agents/light-project.md exists, start from its review profile and
acceptance strategy, then freeze the actual acceptance source.
Core behavior
- Freeze the acceptance baseline and pick a Profile: generic.md, software.md, manuscript.md, agent-skill.md or specification.md.
- Invoke reviewers through
review-loop:generic-reviewfor ordinary artifacts,light-code-reviewfor software diffs, an accepted domain reviewer when one exists. Reviewers only return findings. - Validate findings, direct only bounded in-scope repairs, and ask a fresh, independent Evaluator.
- Issue the verdict when the stopping rules are met.
Entry modes
Use one of three modes; details are in WORKFLOW.md:
init: freeze the acceptance baseline and Profile before review.review: run one bounded evidence, critique, validation, repair, and evaluation round.resume: continue the next unfinished recorded action without rewriting prior evidence.
Handoff and stop
PASS: every frozen criterion has correctly labeled evidence and the fresh Evaluator accepts the baseline.FAIL: a frozen condition is unmet and cannot be resolved through a confirmed, in-scope, bounded repair.BLOCKED: required baseline, authority, environment, evidence, access, or independent context is unavailable; records conflict; or the repair limit stops safe convergence.
Default maximum is three rounds. Record durable state under
.project-review/; see WORKFLOW.md for the exact
state tree and resume rules. When the acceptance source is a tracker ticket,
resolve it after PASS and comment the blocker after FAIL or BLOCKED
(Ticket closeout).
Optional capabilities
- A
FAILwithCause: direction→ after ticket closeout, calllight-redirectwith theverdict.mdpath. Unavailable or declined → the ticket'sFAILcomment is the handoff. .scratch/<feature>/plan.htmlexists → after the verdict and ticket closeout, calllight-html-planto refresh it. Missing page or capability → skip.
References
- WORKFLOW.md — init/review/resume, role packets, state tree.
- acceptance-charter.md — frozen baseline and Profile fields.
- evidence-protocol.md — evidence labels and records.
- finding-schema.md — finding identities, dispositions, registry.
- review-rubric.md — Core checks and Profile dimensions.
- subagent-protocol.md — Critic/Evaluator packets and independence.
- stopping-rules.md — verdict transitions and repair limit.
- migration.md — history only; not needed at runtime.