메뉴
Use when asked for a thorough, deep, or multi-angle review of a branch, PR, or recent changes — or when a diff touches error handling, catch blocks, fallbacks, new types, or test coverage and deserves more than a quick single pass.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Design Apple-style iOS/macOS interfaces following Human Interface Guidelines. Creates HIG-compliant components with SF Symbols, San Francisco typography, and proper accessibility. Supports optional modern effects. Use when designing Apple-style UI, iOS/macOS interfaces, HIG-compliant components, or implementing design system specifications.
Offload mechanical, well-specified coding work (refactors, renames, boilerplate, formatting passes, applying a pattern across many files) to `canaryclaude`, which is Claude Code routed through the user's own CanaryLLM proxy. Cheap and fast enough to fan out — pick a small/fast model like `gemini/gemini-2.5-flash` (or `gemini-2.5-flash-lite`, or local `lmstudio/qwen3.6-27b-mlx`) for the subagent. Use whenever a task is execution-heavy but reasoning-light — code transformations where you already know the answer and just need the typing done. Use even if the user hasn't named canaryclaude: if you're about to do a dozen near-identical edits, or apply a clearly-specified refactor across multiple files, this skill applies. Always background canaryclaude calls and parallelize them — even fast models take 15–60s per task, so it shines when many run concurrently while you do real reasoning work.
Call the CanaryLLM gateway (canaryllm.canarycoders.es) — a multi-provider LLM API for chat completions, image/video/audio/music/dialogue generation, vision (object & face detection, custom model training), conversational AI agents, and drop-in OpenAI/Anthropic/Responses-compatible endpoints. Use this whenever the user wants to hit CanaryLLM / CanaryLLM API / the canary gateway, generate media through it, point an existing tool or SDK at it, or write code/scripts that call any canaryllm.canarycoders.es endpoint — even if they don't paste the spec. Knows the async queue (submit → poll) workflow that the native endpoints require, the Bearer-token auth convention, and which endpoints are synchronous vs queued.
Interface directly with Gem0 (the CouchDB-backed headless CMS / database-as-a-service) over HTTP using API keys — create tables (schemes), modify table schemas, and add/update/query/delete rows. Use this skill whenever the user mentions Gem0, gem0.dev, GemAPI, GemBackend, pillow endpoints, Gem0 schemes/collections/tables, the @gemzero0 SDK, GemSync, or wants any script/integration/agent that reads or writes data in Gem0 — even if they don't say "API" or only say "store this in my CMS".
Control Herdr, a terminal multiplexer for coding agents. Use only when the user explicitly mentions Herdr or asks to use Herdr to inspect or control panes, tabs, workspaces, terminals, commands, or communication with another agent. Do not use merely because a task could benefit from a background terminal, delegation, or parallel work. Requires HERDR_ENV=1.
Use when CLAUDE.md or CLAUDE-*.md memory-bank files need creating, auditing, updating, or syncing with the codebase — after significant code changes, when documented patterns contradict the implementation, at session end to capture learnings, or when context files have grown stale or bloated.
| name | deep-review |
| description | Use when asked for a thorough, deep, or multi-angle review of a branch, PR, or recent changes — or when a diff touches error handling, catch blocks, fallbacks, new types, or test coverage and deserves more than a quick single pass. |
Single entry point for review work. The built-in commands do the core pass; this skill adds specialist lenses that used to live in separate plugins (pr-review-toolkit, code-review, audit, code-simplifier). Those plugins are disabled — if any are still visible, prefer this skill over them.
| Ask | Use |
|---|---|
| Quick correctness pass on working diff | /code-review (low/medium effort) |
| Reuse / simplification / efficiency cleanup | /simplify |
| Security-focused review | /security-review |
| Review a GitHub PR | /review <PR#> |
| Exhaustive multi-agent cloud review | /code-review ultra (user-triggered, billed) |
| Deep local review with specialist lenses | this skill, below |
git diff main... or the PR) and pick only the lenses
that match what changed. Skip lenses with no matching surface.opus
(review and verification are reasoning-heavy — see the model-routing rule
in CLAUDE.md). Each returns findings as file:line — severity (critical/major/minor) — one-sentence defect — failure scenario.silent-failures — catch blocks that swallow errors; broad catches hiding unrelated failures; fallbacks or default values that mask a failed operation; errors logged then execution continues with no user-visible feedback; mock or stub behavior reachable in production paths.
comment-accuracy — comments or docstrings that contradict the code; stale references to renamed symbols; comments that narrate what the next line does or justify the change to a reviewer instead of stating a real constraint.
type-design — invariants left unenforced that the type system could express; leaked internals that should be encapsulated; primitive obsession where a domain type would prevent misuse; impossible states left representable.
test-coverage — new logic without tests; missing edge/error-path cases; tests asserting implementation details instead of behavior; tests that cannot fail.