caph-dev/delegate-grill-with-docs/delegate-grill-with-docs

Delegate Grill With Docs

Workflow

1. Spawn exactly one worker subagent.
2. Fork the current context into it.
3. Leave model and reasoning_effort unset so the subagent inherits the parent agent's model and reasoning strength.
4. The main agent runs the grill-with-docs workflow as the interviewer.
5. Tell the subagent to act as the user's proxy: answer the main agent's questions using the repository, existing docs, and current conversation context.
6. The main agent asks exactly one grill-with-docs question at a time.
7. Send each question to the same subagent, wait for its answer, then evaluate whether the answer is sufficient or another question is needed.
8. Repeat this main-agent-question/subagent-answer loop until the main agent has enough information to stop asking.
9. If the subagent cannot infer an answer from available evidence, it must say what is unknown and what only the real user can decide.
10. After the questioning is complete, the main agent combines the subagent's answers with the target document, updates CONTEXT.md or ADRs when required by the grill-with-docs workflow, and writes the agreed edits back to the original document.
11. Stop after one subagent. Do not fan out or replace it mid-session.

Guardrails

• Trigger only when the user explicitly invokes $delegate-grill-with-docs or names this skill directly.
• Do not auto-run just because the user mentions grill-with-docs.
• Do not ask the real user to answer questions that the subagent can infer from the repo or existing context.
• Do not let the subagent choose the interview agenda; the main agent owns the grill-with-docs questioning loop and final document edits.
• Preserve the project's glossary. If a term conflicts with CONTEXT.md, flag it immediately.
• Update CONTEXT.md inline when a durable term is resolved. Create an ADR only for hard-to-reverse decisions.
• If the user explicitly wants the original live grilling flow, skip delegation and run it directly.