Clarification gates: ask only when a missing fact changes the outcome

Decision in practice

A support ticket says 'cancel the renewal' but lists two contracts for account AC-682. Selecting either contract would cause a different external effect, so the assistant asks which contract ID the user means. In a separate request, the user asks for a short explanation but does not specify bullet count. The assistant can produce a concise paragraph without blocking. A third ticket names contract CT-281, yet the account record shows no current owner; the assistant may explain the process but cannot submit the cancellation until ownership is verified.

Required for action: contract_id and verified account authority.
If two contracts match: ask which ID.
If only output style is unspecified: use concise default and continue.
If owner lookup fails: report review state; do not submit cancellation.
Record: the missing field and why it affects the decision.

Performance and operating cost

An extra clarification turn adds user delay and another model call, so asking about every minor preference can make a system unusable. The gate should be narrow: test whether an unknown field can change the decision or irreversible effect. Checking F required fields is O(F) in application work; external lookups dominate latency. Measure unnecessary questions and wrong assumptions together. Overly aggressive defaults reduce friction but can create wrong account or payment actions, which cost more to repair.

Common Mistakes

• Do not ask for a cosmetic preference when a safe default exists.
• Do not invent a required identifier to avoid asking.
• Do not treat silence as authorization for an external effect.

Connected lessons

• Prompt Engineering
• Output contracts: parse a result and preserve an explicit unknown state
• Human handoff: preserve evidence and the reason for uncertainty
• Zero-shot baseline: measure the task before adding demonstrations
• Prompt envelopes: separate task, evidence, input, and output contract
• Role prompts: use expertise cues without granting authority
• Output budgets: bound length without cutting required facts
• Project: establish a support-triage prompt baseline
• Prompt foundations decisions

Continue with: Spoken confirmations: repeat the risky field, not the whole form.

Continue with: Tutor prompts: start from an observable learning objective.

Continue with: Support prompts: ask only the question that changes the next step.