deep-interview

@deep-interview

Trigger: When scope, intent, or requirements are unclear. Use before @ralplan or @ralph on any non-trivial task.

Depth Profiles

Flag	Threshold	Max Rounds	Use when
--quick	≤ 0.30	5	Minor changes, mostly clear intent
--standard (default)	≤ 0.20	12	Features, refactors
--deep	≤ 0.15	20	Architecture changes, large scope

Protocol

Phase 0 — Preflight

1. Call start_mode with mode=deep-interview and the raw task.
2. Classify: greenfield (new) vs brownfield (existing codebase). For brownfield, read relevant files first — never ask the user for facts you can discover.
3. Call save_artifact type=context, slug=deep-interview-{task-slug}, with initial snapshot:

   Task: ...
   Hypothesis: why the user likely wants this
   Known facts: ...
   Unknowns: ...
   Decision boundaries: ...
   

Phase 1 — Interview Loop

Repeat until ambiguity ≤ threshold OR max rounds reached:

1. Ask ONE question targeting the weakest clarity dimension.

   • Never batch questions. One question per turn.
   • Priority order: Intent → Outcome → Scope → Non-goals → Decision Boundaries → Constraints → Success Criteria
   • For brownfield: use evidence-backed questions ("I see X in Y — should changes follow that pattern?")
   • After each answer: probe assumptions, not just facts. Ask for evidence, counterexamples, or tradeoffs before moving to the next dimension.

2. Score ambiguity after each answer (0.0 = fully clear, 1.0 = fully unclear):

   Greenfield: ambiguity = 1 - (intent×0.30 + outcome×0.25 + scope×0.20 + constraints×0.15 + success×0.10)

   Brownfield: ambiguity = 1 - (intent×0.25 + outcome×0.20 + scope×0.20 + constraints×0.15 + success×0.10 + context×0.10)

3. Show progress each round:

   Round N | Ambiguity: 0.XX → 0.XX | Focus: {dimension}
   

4. Call update_mode_state with phase={dimension} and iteration=N.