Clarifying Questions

Clarifying questions in agentic prompting is a protocol that gates expensive or irreversible operations behind a bounded question round. The model may ask up to N clarifying questions before committing to a workflow, but only when genuinely needed — and after the user responds, the model MUST immediately proceed without re-asking.

This is distinct from “ask for confirmation,” which blocks on a yes/no. Clarifying questions gather disambiguation parameters (definitions, scope, constraints) that genuinely change the output shape.

The Protocol (≤3 Then Launch)

The canonical pattern, as seen in the leaked Anthropic research-instructions prompt, has three rules:

1. Don’t ask for confirmation to launch when the query is already clear, specific, or explicitly requests research. Launch immediately.
2. Pick reasonable defaults when details are unspecified (timeframe, region, examples to include) and note the assumption rather than asking. Only ask when ambiguity would send the work in a completely different direction.
3. After one round of questions, always launch immediately. Never ask clarifying questions twice.

The cap is typically 3 questions. They must be useful, relevant, and genuinely uncertain — no generic, obvious, or inferable questions. Use a numbered list with clear call-to-action.

Worked Examples (from the source)

• “Top 5 fastest-growing startups by revenue growth” → 2 questions (absolute vs. percentage growth; startups-only vs. all companies) because “fastest growing” is ambiguous.
• “Industries that hire non-degreed workers, 2019–2024 trends” → launch immediately (already detailed and specific).
• “Brief this week’s emails” → launch immediately (straightforward, low ambiguity).
• “Taiwan trip plan, October, food/hiking/beaches, $5000/person” → 3 questions (length, interests, budget) because the request is broad enough to fork into different output types.

Comparison with Other Systems

The pattern of “ask, then commit” appears across the corpus but with different caps and triggers:

• Grok Build uses an ask_user_question tool for multiple-choice clarification with recommended-first option ordering, then proceeds.
• Cursor defaults to “discussion-before-code” but uses a solution-only output style (no meta-language).
• Anthropic Tool-Use API describes “discussion-before-code default” as a general product pattern, not a strict instruction count.

The Anthropic research-instructions version is the most prescriptive: explicit cap of 3, explicit “launch after one round” rule, explicit “never ask twice” prohibition.

Implications

A bounded question protocol prevents the “clarification loop” anti-pattern where the model asks 2–3 rounds of questions before acting. The strict cap + immediate-launch-after-answer rule makes the cost predictable: at most one extra round-trip.

When It Goes Wrong

• Asking obvious or inferrable questions (“What date range?” when the user said “this year”)
• Asking more than the cap without noticing
• Re-asking after the user provided answers
• Asking for confirmation disguised as a question (“Do you want me to research X?”)

Relationships

• “references” anthropic-research-instructions — concrete implementation
• “references” communication-style-spectrum — determines whether questions precede explanation or come after solution
• “references” task-state-machine — clarifying questions often gate the first state transition

Sources

• raw/prompts/leaked/anthropic-research-instructions.md