question-protocol

Question protocol

Governs every question the coordinator puts to the founder — concept-interview questions, plan-gate approvals, escalations, attention pings. The installed CLAUDE.md's Execute loop, Escalation, and Watchdogs sections all route founder-facing questions through this protocol rather than defining their own asking style; this skill is what "ask the founder" means in practice.

Every founder-facing question is asked one at a time, never as a batched list or a numbered questionnaire dumped in one message. Each question carries all four of:

1. One line of context: what this blocks / why it's being asked now.
2. The coordinator's own reasoning, briefly — what the agents found, what the actual trade-off is.
3. 2-4 concrete options with a one-line trade-off each, plus a marked recommendation when one exists.
4. The default action if no answer arrives — this must be safe, usually "do nothing yet."

Maintain a question queue when more than one item needs an answer: send the top question, wait for the answer (or an explicit "park this"), then send the next question — never move on to the next question before the current one resolves. That ordering rule is about questions, not work: a pending question blocks only the work that actually depends on its answer — keep working on everything else it doesn't block.

In-session, prefer an AskUserQuestion-style tool if the harness provides one (it renders options natively) with the reasoning folded into the question text. Over a notify channel (Telegram or similar), use the same four-part structure in plain text — see coordinator-kit:comms-register for the notify-channel message format, cadence, and script invocations.