deep-interview

Deep Interview is an intent-first clarification loop that turns vague ideas into execution-ready specifications. It asks one targeted question per round, scores ambiguity across multiple clarity dimensions after each answer, and keeps asking until the score drops below a configurable threshold. The result is a spec that explicitly captures intent, desired outcome, scope, non-goals, decision boundaries, and testable acceptance criteria — everything downstream tools like $ralplan and $autopilot need to work without guessing.

When to use

• The request is broad, ambiguous, or missing concrete acceptance criteria
• You say "deep interview", "interview me", "ask me everything", or "don't assume"
• You want to avoid misaligned implementation caused by underspecified requirements
• You are starting a new feature or project and want a grounded spec before planning begins
• Autopilot or ralph redirects you here because the input was too vague to expand reliably

How to invoke

Natural language triggers: "deep interview", "interview me", "ask me everything", "don't assume".

Explicit slash: $deep-interview

Depth flags (optional):

• --quick — fast pre-PRD pass, up to 5 rounds, threshold 0.30
• --standard — full interview (default), up to 12 rounds, threshold 0.20
• --deep — high-rigor exploration, up to 20 rounds, threshold 0.15

codex
> deep interview "add user authentication to the API"

codex
> $deep-interview --quick "refactor the payments module"

What happens

Deep Interview begins with a preflight context intake, loading or creating a snapshot at .omx/context/{slug}-{timestamp}.md so all later phases share the same grounding. It then classifies the project as greenfield or brownfield and, for brownfield targets, collects relevant codebase facts via explore before asking any questions about internals. Each interview round targets the weakest clarity dimension in priority order — intent and outcome before implementation detail — and displays a weighted ambiguity score alongside the question. After each answer, it applies pressure: it asks for a concrete example, probes hidden assumptions, or forces an explicit tradeoff rather than accepting a surface-level reply. Challenge modes (Contrarian, Simplifier, Ontologist) escalate the pressure at defined thresholds. The loop continues until the ambiguity score meets the threshold, non-goals and decision boundaries are both explicit, and at least one earlier answer has been revisited with a deeper follow-up. Deep Interview then writes a transcript to .omx/interviews/{slug}-{timestamp}.md and an execution-ready spec to .omx/specs/deep-interview-{slug}.md, and offers clear handoff options.

Outputs

• .omx/specs/deep-interview-{slug}.md — execution-ready specification with intent, scope, non-goals, decision boundaries, constraints, and testable acceptance criteria
• .omx/interviews/{slug}-{timestamp}.md — full interview transcript for traceability
• .omx/context/{slug}-{timestamp}.md — grounded context snapshot reusable by downstream skills
• Handoff options: $ralplan, $autopilot, $ralph, $team, or refine further

Related skills

• $ralplan — recommended next step after deep-interview to get architectural consensus on the spec
• $autopilot — full pipeline that accepts the deep-interview spec as its starting brief
• $ralph — persistent execution loop that uses the spec as its completion target