메뉴
Review a code diff before PR or on request. Uses a bounded low-risk fast path, scoped evidence, risk-triggered lenses, and cold-context escalation. NOT browser QA, prepared-plan critique, or external artifact trust checks.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Pull a decorrelated second opinion from a DIFFERENT model on a LOAD-BEARING judgment: a design fork, a pre-commit plan, an irreversible or outward-facing call. Reversible calls just decide. `--panel` = blind critics on a prepared plan. NOT code-diff review (review), NOT mechanical work for Codex (handover), NOT sharpening a fuzzy idea (grill).
Build or fix UI surfaces with a committed point of view: pages, components, layout, typography, motion feel, design review. Triggers: 設計, 做頁面, 做組件, 不好看, 很醜, 突兀, 不協調, 字體, 排版, 樣式, 前端, UI, 截圖, 很死, 死板, 生硬, 沒動效, 沒回饋感, 切換好硬, design, build a page/component, make it look good, typography, feels dead, stiff, lifeless, janky transition, this looks wrong. NOT for backend logic, browser-test verification (use `qa`), root-cause debugging of a broken render (use `debug`), or throwaway variant exploration before a direction is chosen (use `prototype`).
Bounded fresh-context handover for unfinished mechanical build units from an existing plan. A Claude Code or Codex orchestrator may start a fresh Claude or Codex worker and keeps ownership. - /handover [--agent claude|codex] [slug]: sync handoff through `scripts/verbs fresh-run`. - /handover --async [slug]: write a self-contained payload to docs/handoffs/ only; does not spawn Codex or touch git. Use when a locked, file-scoped task benefits from fresh conversation context. NOT for plan writing, closing finished work, PR/ship flow, or exploratory judgment-heavy work (pull a cross-model take with advisor instead).
Focused execution from a concrete outcome to SHIPPED, PAUSED, FAILED, or ABORTED_BY_USER. Adds acceptance-driven loops, bounded review, and delivery evidence beyond native coding behavior. Routes unclear requirements to grill, bugs to debug, UI work to ui, and completed delivery to ship. NOT for hypothetical or planning-only responses.
Audit and reduce a live multi-runtime agent harness after adoption. Use for installed skill/plugin parity, cold-start context, routing overlap, telemetry, or human-attention load. Produces a reversible proposal; it does not mutate the harness.
Adversarial requirement discovery. Ask ONE question at a time, hunting for hidden requirements / unknown unknowns. Drills one angle at a time, then by default runs a structured close that writes a brief (+ executable plan) to docs/. Say "quick" or "don't write files" to stop at a chat-only log instead. Use when the user says "grill me on X", "interrogate this idea", "stress test this scope", "what am I missing", "draft a brief", "structured intake". Skip for tasks where scope is already concrete.
SOC 직업 분류 기준
| name | review |
| description | Review a code diff before PR or on request. Uses a bounded low-risk fast path, scoped evidence, risk-triggered lenses, and cold-context escalation. NOT browser QA, prepared-plan critique, or external artifact trust checks. |
| reads | [{"repo":"**"},{"repo":"AGENTS.md"},{"repo":"CLAUDE.md"},{"skill":"lib/learning-recall.md"},{"skill":"lib/model-anchors.md"},{"cli":"git"}] |
| writes | [{"cli":"stdout"}] |
| forbids | [{"cli":"git push"},{"cli":"gh pr create"}] |
| domain | shared |
| classification | hybrid |
| user-invocable | true |
Native models already perform useful single-pass review. This skill earns its slot through scope provenance, a bounded low-risk path, risk-triggered lenses, cold-context escalation, and findings that survive an evidence gate.
AGENTS.md or CLAUDE.md when present.INTENT GAP and infer only from the diff.Say why, run one grounded correctness pass, and promote when the diff, context, or first pass reveals a higher-risk surface.
A review stays here when scope and intent are explicit, the diff is local and reversible, and it touches no trust boundary, persistent data, concurrency, public interface, generated contract, or production operation. Trace the changed path, match acceptance and branches to tests, then self-refute the assumption most likely to hide a defect.
When that pass finds no candidate finding, concrete coverage gap, scope drift, or risk trigger, return this result and stop:
Review scope: <base>..<head> | <n> files | risk: low
No actionable findings.
Coverage: <verified check>
Self-refute: <assumption and observed result>
Do not load review learnings or model anchors, enumerate lenses, or print empty scope-drift and cold-review fields. Any failed condition promotes the review without restarting scope discovery.
Read lib/learning-recall.md and apply relevant repo learnings. Map changed
surfaces to security, data integrity, concurrency, architecture, or operations
lenses. Medium uses only triggered lenses. High uses every relevant lens plus a
cold review.
Trace changed inputs through code, callers, contracts, tests, and failure handling. Attempt to disprove each candidate. A finding survives only with:
P0 to P3;Exclude style preferences, unreachable speculation, and pre-existing defects outside the diff. Review does not edit code.
Use a cold-context reviewer when the lane is high, the diff exceeds roughly
5K tokens, or a load-bearing conclusion remains disputed. Give it the bound
diff and intent without current conclusions. Read lib/model-anchors.md only
now and select its role. Merge findings by mechanism; disagreement becomes
NEEDS TRACE, not a finding by vote.
Match acceptance and branches to tests, run the narrowest available checks, and
self-refute the highest-risk assumption. Report COVERAGE GAP only for
unproved concrete behavior and SCOPE DRIFT only for changes outside intent.
Conclude with findings, No actionable findings., or BLOCKED.
Review scope: <base>..<head> | <n> files | risk: <lane>
Findings
- [P1] <title> — <file:line>
Trigger: <input/state>
Mechanism: <why it fails and impact>
Direction: <correction>
Coverage: <verified checks or concrete gaps>
Scope drift: <none or entries>
Cold review: <not earned | completed | unavailable>
Self-refute: <assumption and result>