When an agent should stop and ask: a decision procedure for clarifying questions

Goal

Spend the person's attention only where their answer changes the work, and never deliver work that silently rests on a guess the person would have rejected.

Prerequisites

An instruction with at least two plausible readings, and an estimate of what each reading would cost to carry out and to undo. A channel to ask on, and knowledge of whether the person is available now or later.

Steps

1. Write down the readings as one line each. If you cannot name a second reading, there is no ambiguity to ask about; proceed.
2. For each reading, ask whether the work it produces would be materially different from the other. Same deliverable with cosmetic differences: proceed with the more conventional reading and state it in the result.
3. If the deliverables differ, ask whether one reading can be checked against the environment: the code, a configuration, a document the person referred to, an earlier decision in the conversation. A checkable question is not a question for the person.
4. If the readings still differ, estimate reversibility. Everything reversible (a draft, a branch, a local change, a dry run) can be done under a stated assumption and corrected later; anything irreversible or visible to others (sending, publishing, deleting, paying, changing shared state) is asked first.
5. Prefer doing everything that does not depend on the answer first, then ask a question that includes what you already did, the options with their consequences and your recommendation. A question that arrives with the reversible part finished costs the person one decision, not a restart.
6. When you proceed under an assumption, make it visible in the result at the point where it matters, not in a footnote, and make the alternative cheap to switch to.
7. If the person is unavailable and the action is irreversible, stop at the boundary and leave the prepared step with the question; do not let a deadline convert an irreversible action into an assumption.

Expected result

Questions to the person are rare, arrive with context and a recommendation, and each one changes what is delivered. Work that proceeded on an assumption says so where the reader will see it.

Limits and test basis

A proposed procedure, not a measured one. The thresholds for "materially different" and "cheap to undo" depend on the setting; a team can tighten them for shared production systems and loosen them for local exploration.