Most teams teach AI the wrong habit
They tell it to always ask clarifying questions before it starts.
That sounds careful. It usually makes the workflow worse.
The better rule is narrower. Ask only when the answer would materially change the work or reduce meaningful risk. Otherwise proceed, state assumptions, and keep moving.
That conclusion is not just a prompting preference. It lines up with current model guidance, older mixed-initiative HCI research, and newer clarification studies in search and dialogue systems.
The problem with “always ask first”
Blanket clarification rules create a quiet tax on every workflow.
A model that asks before every move slows drafting, interrupts momentum, and pushes routine decisions back onto the human. That matters in knowledge work because many tasks do not need perfect certainty to produce a useful first pass.
If the prompt is “summarize this meeting for the VP,” the model does not need a discovery interview.
If the prompt is “draft three follow-up email variants for this prospect,” the model can probably make progress with a reasonable default tone and one visible assumption.
But if the prompt is “send this to the client,” “change the approval rules,” or “publish this post,” the cost of a wrong assumption is much higher. That is where clarification belongs.
The distinction is simple. Not every ambiguity matters. Some ambiguities change everything.
What the current guidance says
OpenAI’s prompt guidance is the clearest current statement I have seen on this issue. It recommends defining collaboration style up front, including when the assistant asks questions versus makes assumptions, and it explicitly says to prefer making progress when the request is already clear enough. It reserves clarification for cases where missing information would materially change the answer or create meaningful risk.
Source: https://developers.openai.com/api/docs/guides/prompt-guidance
Anthropic’s prompting guidance pushes in the same direction. Give the model context, be direct, and avoid leaving important constraints implicit. Its Skill authoring guidance also matters here. Strong Skills define when they should be used, which conditions change the flow, and what validation loop follows the output. That is a much better operating pattern than “always ask a question first.”
Sources: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices and https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices
Google and Microsoft make the same point from the user side. Better prompts include context, objective, constraints, and fallback behavior. The workflow implication is straightforward. When one of those fields is missing and it matters, ask. When it does not, proceed.
Sources: https://cloud.google.com/gemini/docs/discover/write-prompts and https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/prompt-engineering
Turn one workflow into team infrastructure.
Start with the free Starter Kit if you are still mapping the process. Use the Company-Specific Skill Library Manual when that process needs your tools, data boundaries, review owners, and team language.
What the research says
The older HCI literature gives the design principle.
Clark and Brennan’s grounding work explains why clarification exists in the first place. People need enough shared understanding to coordinate action. Clarification is one way to repair that gap.
Source: https://collablab.northwestern.edu/CollabolabDistro/nucmc/ClarkAndBrennan-GroundingInCommunication-1991.pdf
Horvitz’s work on mixed-initiative interfaces makes the more practical point. Good assistants do not just react to uncertainty. They weigh uncertainty against timing, interruption cost, and the value of resolving the ambiguity before acting.
Sources: https://www.microsoft.com/en-us/research/publication/principles-mixed-initiative-user-interfaces/ and http://erichorvitz.com/chi99horvitz.pdf
McFarlane’s interruption research explains why this matters in daily work. Every clarifying question is an interruption unless it lands at a natural decision point. Too many of them break flow.
Source: https://www.interruptions.net/literature/McFarlane-Interact99-Coordinating.pdf
The newer empirical work supports the same pattern.
Zou et al. found that high-quality clarifying questions improved user performance and satisfaction in search tasks. Low-quality and mid-quality questions did the opposite. In some cases, users were better off without clarification at all.
Source: https://pure.uva.nl/ws/files/169537246/Users_Meet_Clarifying_Questions.pdf
Testoni and Fernández found that model uncertainty does not cleanly match human clarification behavior. That matters because many teams try to solve this by saying “ask whenever you are unsure.” That is incomplete. The useful question is whether asking improves the task enough to justify the interruption.
Source: https://aclanthology.org/2024.eacl-long.16/
The 2025 paper “Clarify When Necessary” makes that tradeoff explicit. The decision is not just whether the prompt is ambiguous. It is whether clarification is likely to improve the outcome enough to justify the extra turn.
Source: https://aclanthology.org/2025.findings-naacl.306/
What to do in a real workflow
The operational fix is not a generic prompting trick. It is a workflow rule.
Treat clarification as a gate.
A simple gate works like this:
- Proceed if the task is clear enough and low-risk.
- Proceed with assumptions if the gap is minor and reversible.
- Ask one narrow question if the missing detail would change the result, create risk, or force significant rework.
- Require approval before external side effects or privileged actions.
That pattern fits the current VibeSec Advisory positioning because governed AI workflows need explicit review points, not vague advice to “be careful.” A good workflow makes the default path obvious.
It also fits how strong Skills should be written.
Instead of telling a model “ask clarifying questions if anything is ambiguous,” write the policy directly:
- Ask when audience, scope, evidence standard, or approval status materially changes the output.
- Use the default template when those details do not matter.
- Before any external action, confirm or escalate.
- If a safe partial step exists, do that first.
That keeps the Skill usable. It also makes the behavior reviewable.
A better prompting habit
There is still a useful prompting habit here. It is just more precise than most people make it.
Do not end every prompt with “ask clarifying questions before you start.”
End it with a rule like this:
“If a missing detail would materially change the answer or create meaningful risk, ask one narrow clarifying question. Otherwise proceed and state your assumptions.”
That one sentence does more than slow the model down. It teaches judgment.
For teams redesigning AI-assisted work, that is the real goal. Not maximum caution. Not maximum speed. Controlled progress with visible assumptions, visible gates, and fewer bad surprises.