Unpack vague requirements before building

Interview user to unpack underspecified requirements before planning

Pre-spec phase where guessing wrong is still free—unpack vague asks like 'build me a dashboard' before committing to architecture.

• · User's initial request (vague or conventional)
• · Your hypothesis about what they actually want (confidence 0-100%)

• · Series of 1-question-at-a-time interview prompts with your guess attached
• · User reactions that validate or refute each hypothesis
• · Updated hypothesis confidence after each answer
• · Final hypothesis ready for spec/plan phase

• · None (pure conversational)

• · Live, responsive user (not CI/scheduled/non-interactive)
• · Request is missing at least one of: WHO, WHY, SUCCESS metric, CONSTRAINT
• · Confidence on initial hypothesis < 80%

• · Invoking in non-interactive context (CI, scheduled, /loop) → hangs waiting for user
• · User is agreeable and doesn't push back → missed signal, locked in wrong assumption
• · Asking in batches instead of 1 question at a time → loses the diagnostic power
• · Not attaching your guess → user generates answer from scratch instead of reacting fast to wrongness

• · Stop condition: 'Do not invoke in non-interactive contexts'
• · Explicit when NOT to use: mechanical edits, unambiguous requests, pure info (no decision risk)
• · Hypothesis format: 1 sentence + 0-100% confidence number
• · Question format: 'Q: ... GUESS: ... [reasoning]'
• · One-at-a-time rule with reasoning (prevents skim-reading, lets user react fast)