Elicit

Bound the blocker

State goal, blocker, decision, evidence, and ownership. Check the repository, tools, supplied policy, conventions, and authority before asking. Ask facts only when they cannot be derived; never turn an evidence gap into a preference.

One-question protocol

Ask exactly one concise question containing the smallest missing user-owned fact, choice, permission, or authority needed to change the action. Do not bundle questions or gather generic or background preferences; when Elicit was invoked for a material preference, ask only for the one specific preference whose answer changes the action. Do not choose user-owned policy, risk, scope, permission, or operational rules; present known materially distinct choices. When existing evidence genuinely supports a default, include a concise evidence-backed recommendation and important consequence in that question. For pure preference or authority choices, ask neutrally; never manufacture a recommendation.

Re-evaluate and stop

After every answer, re-evaluate the blocker and stop immediately when one justified next move exists:

• return the next move to the invoking owner when unblocked; route through Think only when Think invoked Elicit;
• return to bounded exploration for a derivable gap;
• return to owner for agent-owned judgment, including explicit user delegation (“you decide”/“do what seems best”) within its domain and safety authority; treat as returned authority, not ambiguity;
• ask one further minimum question only for a distinct material user-owned blocker that changes the action;
• if absent, genuinely ambiguous, or insufficient, report the exact blocker and missing authority without looping.