Two explicit-mode Codex skills for heavier planning and optimization workflows.
These skills are intentionally opt-in. They do not replace Superpowers, ECC, or a project's own AGENTS.md; they provide named modes you can invoke when a task needs more structure than the default workflow.
Use when you explicitly want a main Planner to coordinate a complex task, compare approaches, split work, delegate bounded subagent tasks, integrate results, and verify before completion.
Best for:
- Multi-module implementation plans.
- Tasks that need research, implementation, and review split apart.
- Work where ownership, acceptance criteria, and verification need to be explicit.
Not for:
- Small edits.
- Single-step fixes.
- Tasks where direct implementation is faster and safer.
Use when you explicitly want a council-style optimization process with real subagents. The council must run 3-6 debate rounds and include perspectives for architecture, efficiency, user value, and Red Team risk review.
Best for:
- Product, workflow, prompt, paper, or system optimization.
- Decisions with competing goals such as maintainability, cost, quality, and user value.
- High-impact changes that need structured challenge before execution.
Not for:
- Routine bug fixes.
- Simple performance suggestions.
- Situations where subagents are unavailable and you do not want a simulated fallback.
Copy the skill folders into your Codex skills directory:
mkdir -p ~/.codex/skills
cp -R skills/planner-coordinator ~/.codex/skills/
cp -R skills/optimization-council ~/.codex/skills/Restart Codex so the new skill metadata is loaded.
Use planner-coordinator mode to plan this multi-module refactor before implementation.
Use optimization-council mode to evaluate this product workflow. Run the minimum 3 council rounds.
启动优化议会,针对这个论文工作流做多角色优化,必须启动 subagents。
- Explicit mode only: these skills should trigger when requested by name or by a clearly matching user request.
- Proportional process: do not use the Planner for tiny tasks.
- Real council:
optimization-councilrequires subagents unless the user explicitly accepts a simulated fallback. - Verification matters: debate and planning do not replace project tests, lint, build, review, or domain-specific gates.
MIT