메뉴
Use when user wants to delegate a task to OpenAI Codex CLI. Triggers on "ask codex", "have codex", "codex:", "send to codex", "let codex", "run codex". Delegates via subagent to keep main context clean.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Create and manage dedicated Codex lane threads with goals, acceptance criteria, PR babysitting, review gates, and wrap-up. Use when work should be split across Codex threads, or when an existing diff needs a manager thread to review, PR, babysit, and ship it.
Use when the user wants to brainstorm, stress-test, sharpen a plan, explore options, decide what to build, or says "grill me", "/grill-me", "help me think through", "sharpen this idea", or "what should I build".
Pre-merge review that runs PR quality, tech debt, security, regression, and performance analysis in parallel via general-purpose agents, aggregates findings into a unified prioritized report, then auto-fixes mechanical issues. Use when the user says "final review", "pre-merge review", "run all reviews", or wants a comprehensive check before merging. Defaults to all reviewers; accepts args to run a subset (e.g., `/final-review security techdebt`).
Deep code audit via Codex with full-access sandbox and validation gate. Triggers on "codex audit", "deep review", "audit with codex", "thorough code review", "/codex-audit". Unlike /codex, this gives Codex full access to run tests and explore, then validates every finding before presenting.
Multi-round iterative review via Codex CLI with session resumption — Codex reviews a plan, design, or diff, returns a verdict (APPROVED/REVISE), and Claude fixes issues and resubmits until Codex approves. Codex remembers all prior feedback across rounds via session resume, so it verifies fixes rather than re-discovering the same issues. MUST use this skill when the user wants Codex to review a plan, design doc, architecture proposal, or diff iteratively with multiple rounds of feedback. Trigger phrases include "codex iterative review", "iterative review", "review my plan with codex", "have codex review this iteratively", "multi-round review", "review until approved", "get codex to tear it apart", "codex review and verify fixes", "not just a single pass", "keep reviewing until it's solid", or "/codex-iterative". Also trigger when the user asks for a Codex review where they want verification that fixes address prior feedback, session persistence across review rounds, or accumulated context. Do NOT trigger for q
Use when you have 2+ independent tasks to run concurrently without a formal plan. Triggers on "parallel research", "parallel subagents", "explore in parallel", "investigate multiple", "run tests in parallel". NOT for plan execution (use subagent-driven-development instead)
SOC 직업 분류 기준
| name | codex |
| description | Use when user wants to delegate a task to OpenAI Codex CLI. Triggers on "ask codex", "have codex", "codex:", "send to codex", "let codex", "run codex". Delegates via subagent to keep main context clean. |
Delegate tasks to OpenAI's Codex CLI via a subagent. Codex runs non-interactively and results are summarized back.
Is this a code review request?
│
┌────┴────┐
yes no
│ │
▼ ▼
codex codex exec
review --sandbox read-only
Dispatch a single Bash subagent (Task tool) with the appropriate command. This keeps all Codex output isolated from the main context window.
Use codex review which has specialized review behavior:
Task tool:
subagent_type: Bash
model: haiku
description: "Codex review"
prompt: |
Run this command and return the full output:
codex review {FLAGS} "{INSTRUCTIONS}" -o /tmp/codex-out.txt 2>/dev/null; cat /tmp/codex-out.txt
Flags:
- --uncommitted → review staged/unstaged/untracked changes
- --base BRANCH → review changes against a base branch
- --commit SHA → review a specific commit
Default to --uncommitted if user didn't specify.
Return the complete review output.
Use codex exec in read-only sandbox:
Task tool:
subagent_type: Bash
model: haiku
description: "Codex: {brief task}"
prompt: |
Run this command and return the full output:
codex exec "{USER_PROMPT}" --sandbox read-only -C "{WORKING_DIR}" -o /tmp/codex-out.txt 2>/dev/null; cat /tmp/codex-out.txt
Return the complete output.
User must explicitly ask for write access. Use --sandbox workspace-write instead of read-only:
codex exec "{PROMPT}" --sandbox workspace-write -C "{WORKING_DIR}" -o /tmp/codex-out.txt 2>/dev/null; cat /tmp/codex-out.txt
| User Says | Route To | Flags |
|---|---|---|
| "ask codex to review" | codex review | --uncommitted |
| "codex review against main" | codex review | --base main |
| "codex review commit abc123" | codex review | --commit abc123 |
| "ask codex to analyze/explain/plan" | codex exec | --sandbox read-only |
| "have codex fix/refactor/implement" | codex exec | --sandbox workspace-write |
| Mistake | Fix |
|---|---|
| Running codex in main context | Always use a Bash subagent via Task tool |
Forgetting 2>/dev/null | Codex thinking tokens flood context without this |
Forgetting -o flag | Without it, output mixes with metadata noise |
| Using interactive mode | Always use codex exec or codex review, never bare codex |
| Hardcoding model | Let user's codex config choose the model |