/consult-codex skill for Claude Code.

SKILL.md

name	description
consult-codex	Consult OpenAI Codex CLI for a second opinion or alternative perspective. Triggers on "consult codex", "ask codex", "check with codex", "codex opinion", "what does codex think".

Consult Codex

Get a second opinion from OpenAI's Codex CLI on the current topic, question, or task. Use the output to inform your next action based on the user's request.

When This Triggers

• User explicitly says "consult codex", "ask codex", or similar
• User wants a second opinion on debugging, architecture, or implementation
• User instructs you to consider Codex's perspective before planning or acting

Core Workflow

1. Formulate the prompt — Distill the current conversation context into a clear, actionable prompt for Codex. Include:

   • The specific question, bug, or task at hand
   • Relevant file paths or code snippets (keep it concise)
   • What kind of answer is needed (explanation, fix, alternatives, review)
   • If the user provided an explicit message for Codex, use that instead

2. Run Codex — Execute via Bash using the non-interactive exec mode:

   codex exec --sandbox read-only "YOUR PROMPT HERE"

   • --sandbox read-only restricts Codex to read-only file access (no modifications)
   • Codex runs in the current working directory and can see project files
   • stdout contains the final answer; stderr has progress (ignore stderr)
   • For longer prompts, use a heredoc:

     codex exec --sandbox read-only "$(cat <<'EOF'
     Your multi-line prompt here.
     Include context, file paths, and what you need.
     EOF
     )"

3. Interpret the output — Based on the user's request:

   • Debugging: Compare Codex's diagnosis with your own, highlight agreements and new insights
   • Planning: Incorporate Codex's suggestions into your plan, noting where approaches differ
   • Code review: Present Codex's feedback alongside your own assessment
   • General question: Summarize Codex's answer and add your own perspective if relevant

4. Act on the result — Follow the user's intent:

   • If user asked to "consult then implement" → use the combined insights to write code
   • If user asked for a second opinion → present both perspectives and let the user decide
   • If user asked to debug with Codex → synthesize findings into a resolution

Prompting Tips

• Be specific about scope: "Look at src/features/tags/queries.ts and explain why the cache isn't invalidating after mutation"
• Include constraints: "This project uses TanStack Query v5 and Zustand — suggest solutions using these libraries"
• Ask for concrete output: "Suggest a fix" rather than "What do you think about this code"
• Keep prompts under ~500 words for best results

Do

• Always use codex exec --sandbox read-only for read-only consultation (no file modifications)
• Summarize Codex's response concisely for the user
• Clearly distinguish between Codex's suggestions and your own
• Mention when you and Codex agree — convergent opinions build confidence

Do NOT

• Let Codex modify files — always use --sandbox read-only
• Blindly follow Codex output — verify against project conventions in CLAUDE.md
• Run Codex for trivial questions you can answer directly
• Pass sensitive data (API keys, credentials) in the prompt