Interview

Turns an underspecified request into a confirmed, thought-through scope before a single line is planned or written. It restates and diffs the request, researches the code so its questions are specific, surfaces the unknowns and non-goals, interviews the user only where research falls short, and emits a brief.md that planning consumes.

This SKILL.md is a thin index. Phase procedures live in rules/*.md and load on demand. A full worked run lives in references/interview-example.md.

Contents

• When it runs
• Workflow
• Required reading by phase
• Core principles
• Anti-patterns
• Definition of done

---

When it runs

$ARGUMENTS is the request to align on (free text). If $ARGUMENTS is empty, use the current conversation's active request; if there is none, ask what to scope — never interview against a guessed request.

This skill is adaptive. When the request is already crisp and research answers the open questions, it asks nothing, writes a one-line pass note plus the brief, and returns ready. It interrupts only for a genuine, load-bearing unknown — see rules/readiness-gate.md.

Relation to aw and ideate (precedence)

• Autonomous-implementation asks ("implement X autonomously", "ship this", /aw) are owned by aw, which runs this interview inside its Phase 0 (default-on for Full tier — see rules/aw-integration.md) — do not also fire standalone; let aw drive.
• Standalone /interview is for the explicit "help me scope this first" moment on non-aw or ambiguous work.
• For solution options, not scope, use /ideate — it diverges and scores; this skill converges on the one scope.

Workflow

Five phases, each with a gate. Do not advance until the gate passes.

Required reading by phase

Load on demand — do not preload.

Core principles

1. Research before asking. A question the codebase already answers is noise. Ground every question in what the code shows so it is specific ("reuse PrStatusCache or add a cache?"), not generic ("what do you want?"). See rules/research-grounding.md.
2. Interrupt only for load-bearing unknowns. Classify each unknown blocking vs advisory. Ask about blocking ones; proceed on advisory ones and record the assumption. Silence is the correct output for a crisp request. See rules/readiness-gate.md.
3. Batch the questions. One prioritized round via AskUserQuestion, two at most. Never drip questions one at a time. See rules/question-design.md.
4. Consultative, not stenographic. Capture intent and pressure-test completeness: non-goals, edge cases, success criteria, constraints. This is the "is it thought-through" half. See rules/consultative-completeness.md.
5. Converge, don't diverge. This skill pins down the one thing to build. It hands divergent option-generation to /ideate, adversarial plan review to /critical, and readiness scoring to /confidence. Do not duplicate them.
6. The brief is the deliverable. A confirmed, self-contained brief.md plus a readiness verdict — nothing more. It writes no product code. See rules/brief-artifact.md.

Anti-patterns

• Asking a question the code answers. Fix: research first (Phase 1).
• Dripping questions one at a time across turns. Fix: batch into one AskUserQuestion round.
• Interviewing a request that is already crisp. Fix: honor the adaptive pass — write the brief, return ready, ask nothing.
• Generating solution options. Fix: that is /ideate; this skill converges on scope.
• Adversarially red-teaming an implementation plan. Fix: that is /critical; run it after a plan exists.
• Producing a scored gate verdict. Fix: that is /confidence; this skill elicits the inputs a score needs.
• Writing product code or a plan. Fix: stop at brief.md; hand off to planning.

Definition of done

• Request restated in own words; every delta from the user's words surfaced (Phase 0).
• Affected code areas, prior art, and constraints located (Phase 1).
• Every unknown enumerated and classified blocking vs advisory (Phase 2).
• Blocking + high-value unknowns resolved via ≤ 2 batched rounds, or a clean adaptive pass logged (Phase 3).
• brief.md written to .agent/{branch}/brief.md (re-run updates it in place).
• Readiness verdict set: ready | ready-with-assumptions | blocked.
• One-line summary delivered naming the verdict and pointing at the brief.