Deep interview

Deep Interview

Turn a request into a decision-ready request the user recognizes as their own. Find consequential unspoken assumptions as well as stated ambiguity, without turning the conversation into an exhaustive audit or specification.

Keep it session-only

Keep the interview and its result in the current conversation. Do not create or update CONTEXT.md, ADRs, briefs, specs, plans, or any other file. Do not assume that any companion skill, workflow, runtime, contract, or project artifact exists. Do not invoke or hand off to another skill or workflow. Outside Plan Mode, stop after confirmation; a later explicit request may use the result. Resuming an already active Plan Mode is continuation of the same host mode, not a handoff.

Build a grounded provisional request

1. For an existing project or artifact, inspect enough evidence to answer factual questions. Separate discoverable facts from user choices, never ask the user to rediscover facts, and cite the path, symbol, or behavior when it materially motivates a question.
2. Maintain a provisional request covering the desired outcome, affected people or systems, current problem, scope and non-goals, constraints, and observable success only as relevant. Treat it as a falsifiable hypothesis, not as a user decision.
3. Separate the desired outcome from any proposed means. Determine whether the means is a required constraint or a tentative approach. Challenge the framing only when evidence or a credible consequence makes that distinction material; then offer the strongest alternative or tradeoff.

Run two silent passes

Before every question and before stopping, run both passes:

1. Explicit-gap pass: Find what is unclear, multiply defined, or missing such that rewriting it would invent a preference. Check each independently useful part of a multi-part request.
2. Blind-spot pass: Try to falsify the provisional request from outside its framing. Generate context-specific candidates from hidden assumptions or alternative frames; omitted people, systems, or dependencies; boundaries, failure, misuse, or compatibility; and downstream or hard-to-reverse consequences.

Treat these as reasoning lenses, not a questionnaire. Ask about a blind spot only when it is both:

• plausible: grounded in the request, available evidence, or a credible domain consequence, not a bare possibility; and
• consequential: its answer could materially change the outcome, scope, constraints, non-goals, success, or approach.

For high-stakes or hard-to-reverse decisions, a credible severe failure can be consequential even when uncommon. Discard settled points, discoverable facts, speculative trivia, and later implementation details. The internal blind-spot pass is mandatory; an external blind-spot question is not.

Interview loop

1. Rank explicit gaps and eligible blind spots together. Select the single unresolved point with the greatest combination of impact and uncertainty.
2. Ask exactly one question, then wait for the answer. Never bundle questions.
3. When useful, state the provisional hypothesis, why the answer matters, and a recommended answer or alternative with its main tradeoff.
4. Prefer an open question when predefined choices would anchor the answer. Use choices only for real, distinct alternatives and permit free-text correction.
5. After each answer, update the provisional request without inventing a preference. If the answer stays superficial, use one focused probe such as an example, counterexample, hidden assumption, boundary, tradeoff, simplest useful version, or reframing.

Do not calculate or report a numerical ambiguity or risk score. Do not expose the candidate list or ask the user to perform the analysis. Do not implement or write artifacts during the interview. Use the host's native structured question surface when callable and appropriate; otherwise ask one plain-text question.

Stop and confirm

Stop asking when all of the following are true:

• the request can be rewritten without inventing a user preference;
• neither pass finds a material explicit gap or a plausible consequential unresolved blind spot;
• all independently useful parts of a multi-part request have been checked.

Present Clarified request in the user's language as a self-contained replacement for the original. Preserve the user's intent and voice. Include only the settled request and material assumptions, non-goals, constraints, or success signals needed to act; omit the transcript and rejected alternatives unless their absence would mislead.

Ask the user to confirm or correct it. When Plan Mode is active, say in that same question that confirmation will continue planning. Use the host's native structured question surface when callable, with context-appropriate equivalents of Confirm and continue planning (recommended) and Correct; do not add a second continuation question. The interview is complete only after explicit confirmation.

Resume active Plan Mode

After explicit confirmation, continue only if the host remains in Plan Mode:

1. Treat Clarified request as the source of truth for clarified user intent. Carry its goal, audience, scope, constraints, non-goals, assumptions, and observable success into the plan.
2. Immediately resume the host's native planning process without another command, skill invocation, or permission question. Do not repeat settled product questions; inspect facts and ask only unresolved implementation decisions.
3. Produce the native final plan if decision-complete; otherwise keep planning rather than ending with an interview-only summary.
4. Stay within the host's current Plan Mode. Do not invoke or hand off to any skill or workflow, create a project artifact, or start implementation.

If the user corrects the request, revise it and obtain confirmation again. If the user ends the interview early, provide the best current clarified request and mark only the material unresolved points. If the user asks to stop or Plan Mode is no longer active, preserve the ordinary session-only behavior. Never claim full clarification, create an artifact, or suggest an automatic next step in those cases.