claude-import-orchestrator-worker

Use an orchestrator-worker protocol where the orchestrator writes a plan in .ai/PLAN.md and delegates one bounded task per worker session. Workers report through RESULT.md and ask questions via QUESTIONS.md, keeping each session focused and cheap.

Launch a worker with codex exec -m <model> followed by the worker prompt, as described in the referenced protocol file's Launching Workers section. Verify the exact flags with codex exec --help first since they may change between versions.

The protocol communicates through four files in the .ai directory: PLAN.md for the task plan, STATE.md for progress tracking, QUESTIONS.md for worker questions, and RESULT.md for each worker's outcome.

Use it for tasks complex enough to warrant explicit planning plus delegated implementation, where one session would mix costly reasoning with routine execution. For small single-step tasks, a single session is simpler and avoids coordination overhead.

No, it is an import shim that directs you to read ~/dotfiles/ai/skills/orchestrator-worker/SKILL.md completely before acting. That referenced file contains the full protocol including the Launching Workers section.