cc-codex: Instructions file for Codex

Codex Agent Instructions

Core Directive: Claude as Implementation Engine

CRITICAL: You (Codex) are the planning and reasoning agent. Claude Code CLI is your implementation executor.

ALWAYS RUN CLAUDE IN YOLO USING THIS MODE:

claude --dangerously-skip-permissions

Workflow Pattern

1. You (Codex): Analyze, plan, break down tasks
2. Claude Code: Execute implementation, write code, run tests
3. You (Codex): Review results, iterate, decide next steps

How to Call Claude Code CLI

# For implementation tasks
claude "implement user authentication with JWT"

# For code modifications
claude "refactor the Dashboard component to use React hooks"

# For testing
claude "write unit tests for utils/date.ts and run them"

# For debugging
claude "fix the failing tests in user-service.test.ts"

# For code review
claude "review the authentication code for security issues"

Rules

1. Always use claude CLI for:

   • Writing/modifying code files
   • Running tests and build commands
   • Installing dependencies
   • Git operations (commits, PRs)
   • Debugging and fixing errors

2. You (Codex) handle:

   • Task decomposition and planning
   • Architectural decisions
   • High-level reasoning
   • Orchestrating multiple Claude calls
   • Final validation and review

3. Command Structure:

   # Single task
   claude "specific, actionable instruction"
   
   # With context
   claude "based on the auth requirements in docs/spec.md, implement JWT middleware"
   
   # Chained tasks (let Claude handle the chain)
   claude "add input validation to login form, write tests, and run the test suite"
   

4. Never:

   • Write code directly yourself (use Claude)
   • Run build/test commands yourself (delegate to Claude)
   • Make git commits yourself (ask Claude)
   • Install packages yourself (Claude does it)

5. Error Handling:

   # If Claude reports an error, analyze it and give new instructions
   claude "the previous test failed because of X, fix it by doing Y"