Interview

Interview: Requirement Clarification

Turn a vague request into a spec the user actually wants built, through a hypothesis-driven interview.

The request the user gives is a map; the codebase and its real constraints are the territory. Every gap between them is an unknown that would otherwise get filled with a guess. Instructions that are too specific get followed even when a pivot is better; instructions that are too vague get patched with industry defaults that may not fit. The interview escapes both failure modes: instead of guessing, put the decision points in front of the user — including the ones they never thought to raise. A few minutes of questions is the cheapest way to find out what nobody knew, before it gets expensive to fix.

When to Use

• Ambiguous feature requests ("add a login feature")
• Incomplete bug reports ("the export is broken")
• Underspecified tasks ("make the app faster")

For strategy/planning blind spot analysis, use unknown. For content-vs-form reframing, use metamedium. If the user's uncertainty is "I'll know it when I see it" (visual/design taste), offer 2-3 quick prototypes or variants instead of more questions — interviews resolve ambiguities that words can settle.

Invocation

Parse input:

• REQUIREMENT: requirement text (required)
• --max-iterations N: max question rounds (default: 3)

Initialization

1. Create .claude/clarify-interview.local.md at the project root (the current working directory). The Stop hook resolves the same path from the hook cwd.

---
iteration: 1
max_iterations: [MAX_ITERATIONS]
original_requirement: "[REQUIREMENT]"
started_at: "[ISO 8601 timestamp, e.g. 2026-07-10T14:30:00+09:00]"
---

## Original Requirement
"[REQUIREMENT]"

## Clarification Progress
(Append one block per round — see Interview Rounds step 4)

started_at is required: the Stop hook has no session id at skill time, so it binds the loop to the current session on its first fire (adds session_id itself) and uses started_at for a 2-hour TTL. Do not add session_id yourself.

2. Create the companion guard file .claude/clarify-interview.guard in the same directory, containing one line: [iteration] [max_iterations] _ [current unix epoch] (e.g. 1 3 _ 1783728000 — get the epoch with date +%s). This is the hook's tamper-evidence: it snapshots the loop position so tag validation survives even if the state file goes missing; the epoch anchors its 2-hour TTL to creation time, and the hook rejects guards without one. Never delete it — the hook consumes it.

3. Confirm activation:

Interview clarification loop activated!

Original Requirement: "[REQUIREMENT]"
Max Iterations: [MAX_ITERATIONS]

Say "cancel" or "stop" anytime to end the loop.

Protocol

Phase 1: Walk the Territory

Before designing any question, spend a few minutes in the actual project: the files the request touches, recent related changes, README/config that constrains the answer. Two reasons:

• Asking the user something the codebase already answers wastes their attention and your question budget.
• Questions grounded in what you found ("this route was rewritten last month — broken since then?") get far better answers than template questions.

If there is no project context (greenfield conversation), skip the sweep and ground questions in the request's own domain instead.

When the project carries cheap, side-effect-free checks (existing test suite, typecheck, lint), run them during the walk — a currently-red test or broken script is a territory fact the user may not know, and it often becomes a Success Criterion or a scope question.

Phase 2: Diagnose Ambiguity

List candidate unknowns across: scope, behavior, data/interface, constraints/priority, success criteria. An ambiguity is material when different plausible answers lead to different implementations — if every answer lands in the same code, don't ask.

Then look past the user's own framing: what has this user not considered at all? Failure and abuse paths, volume/rate extremes, lifecycle after the happy path (what happens after the third retry fails?), who else is affected. Plan at least one question or option per interview that surfaces such a blind spot — the mark of a good interview is the user saying "oh right, that too." If the territory suggests the request itself may be solving the wrong problem, raise that as a question, not a silent redirect.

Phase 3: Interview Rounds

Use AskUserQuestion. This section is the single source of truth for question shaping — the Stop hook only re-injects a pointer back here, it does not restate the rules.

Runtime note — capability ladder (every rule below applies at each step):

• Claude Code: AskUserQuestion — up to 4 questions per call, 2–4 options each, multiSelect supported.
• Codex: request_user_input where available (it is single-select and may be absent outside plan mode; its app-server contract documents 1–3 questions per call, so keep a round to ≤3 questions there). Ask a would-be multiSelect question as plain text ("해당하는 것 모두"), and when the tool itself is unavailable, drop to the plain-text form below.
• Plain text (any runtime): present the same batched questions as numbered lists — 1) 옵션 / 2) 옵션 / 3) 기타: 직접 설명 — and wait for one reply covering all questions; say explicitly that multiple picks are welcome where choices aren't exclusive. Before ending a turn that prints plain-text questions, set awaiting_user: "true" in the state-file frontmatter (add the key if absent) — the Stop hook lets exactly that turn end so the user can actually reply, then flips the flag back; without it the hook would re-inject the loop over your pending questions.

1. Prioritize by leverage. Lead with questions whose answers change the architecture or scope (who is it for, what triggers it, what's in/out); leave wording/naming details out entirely. When the requester's authority looks limited (proxy signals, ops role, "회사 시스템"), front-load the axes any operator can answer — who is it for, which screens/data, today's workflow — and defer owner-policy axes; a round spent on questions the requester must relay to someone else harvests only deferrals.
2. Shape each question: 2–4 options, each a plausible hypothesis of intent that would genuinely change what gets built; one ambiguity axis per question; neutral framing (don't make one option "obviously right" by wording); concrete, implementation-relevant language, in the conversation's language.
   • Bad: What kind of login do you want?
   • Good: OAuth / Email+Password / SSO / Magic link
   • For requests that change an existing system (migrate, convert, adopt, rewrite): include adoption strategy among the hypotheses — incremental, big-bang, tooling/check-only — not just end-state variants, and give the scope question itself at least one trigger-policy hypothesis (e.g. 신규 파일만 + 기존 파일은 손댈 때 전환). End-state-only options force users to smuggle "gradually, please" through Other. On a running system, never label the most invasive end-state (Recommended) unless the territory gave you a concrete reason.
   • When you have a sensible default, put it first labeled (Recommended). Use multiSelect: true when choices aren't mutually exclusive. A delegation option that names its consequence (잘 모르겠어요 — 추천대로 (X로 가정)) beats a bare "모르겠음": the user sees what deferring commits them to.
3. Batch independent questions — up to 4 per round, one AskUserQuestion call. If a question's options depend on another question's answer, defer it to the next round: options designed blind degrade into guesses.
4. After each round, append the outcomes to the state file under ## Clarification Progress. Keep iteration equal to the round you are currently asking — bump it as you start a round, not after the last one — and mirror every iteration bump into the first number of .claude/clarify-interview.guard (never lower it). Sessions get compacted and interrupted; the state file is what lets the loop resume without re-asking, and the guard is what lets the hook validate the terminal tag even if the state file disappears.

### Round N
- [axis]: [outcome] — decided | assumed (recommended default) | still open

5. Handle "I don't know" / free-text answers: when the user defers ("잘 모르겠어요, 추천해주세요"), adopt your recommended option and record it as assumed, not decided. A custom "Other" answer is an explicit user statement — record it as decided; it usually opens a fresh hypothesis, so follow up next round only if material. When the answer defers to an absent authority, the requester is a proxy: record your recommended default as assumed — pending [owner]'s confirmation, or put the item in Still Open naming the owner and how it gets resolved. Authority deferral is semantic, not phrasal — "그건 대표님이 정하셔야 해요", "제가 정할 운영 방식은 아닌 것 같아요", "팀에 물어봐야 해요" all count. If the owner is unnamed and a question slot remains, ask who decides; otherwise Still Open reads owner 미확인. Never present a deferred axis as decided, and never let it become a plain assumption without the pending-owner marker.
   • Greenfield caveat: if the project is empty but the request implies an existing system elsewhere, ask where this will integrate — and specifically who or what can write to the target system (read-only access, a gatekeeper dev, an API): the write path usually decides the deliverable's form (hand-off file vs direct integration). An unknown integration point blocks the build, so it belongs in the interview, not in Still Open.
6. Offer an exit honestly. When remaining ambiguities are optional depth, offer the exit as its own 2-option control question (계속 다듬기 / 여기까지 — 현재 이해로 진행, localized) — never as an extra option stuffed into a domain question, which merges two axes into one question. If the round's payload already needs all 4 slots for real hypotheses, skip the exit question and rely on Phase 4 judgment instead. When material ambiguities clearly remain, don't offer a premature exit; when none remain, don't ask another round at all — summarize.

Phase 4: Completion Check

Summarize when any of these holds:

• No material ambiguity remains (your judgment — the common case)
• User selects the clarification-complete option
• User signals cancellation (see Cancellation)
• Final round's answers are in and max_iterations is reached

Hitting max_iterations with material axes still unasked is triage, not completion: spend the final round on correctness-critical axes first. Each unasked material axis goes under Still Open labeled un-asked (e.g. 미질문 — 라운드 예산 소진) with how it gets resolved — not as an assumption carrying a concrete default you invented, and never as something the user delegated. A capped interview ends with <promise>CLARIFICATION CAPPED</promise> instead of the COMPLETE tag (Phase 5); ending this way is a legitimate, expected exit — emit the summary without hesitation.

A fully-consumed budget always ends CAPPED — even when every axis you thought of was asked and resolved. You can only triage axes you conceived of; a burned budget left no room to probe for the ones you didn't, so "no material axis remains" is a claim you cannot make from a spent budget (that exact claim has been measured false: interviews that declared it were missing a spec-changing axis the requester never volunteered). COMPLETE is reserved for interviews that end themselves with budget to spare. In the all-axes-resolved capped case, Still Open carries one honest line — 라운드 예산 소진 — 미탐지 축 잔존 가능성, 구현 첫 리뷰에서 재확인 — instead of invented un-asked topics.

Cancellation

If the user signals they want to cancel or stop the loop (e.g. "cancel", "stop", "그만", "중단"), delete .claude/clarify-interview.local.md and .claude/clarify-interview.guard, then end the loop without a summary.

Phase 5: Before/After Summary

Every line of the After spec must trace to a user answer, a fact you verified in the territory, or a listed assumption — a spec line with no source is fabrication. Before emitting the promise, audit the After spec claim by claim — parentheticals included (volumes, thresholds, counts, quotes) and motivation/state claims ("팀 보고가 막혀 있어", "X가 매주 쓴다"-류) — and write the audit down: append a ### Claim audit block under ## Clarification Progress in the state file, mapping each factual claim to its source (R2.Q1 답변 | territory src/db.js:12 | 가정 #2). A claim you cannot give a source line gets rewritten as an assumption or dropped. The label or description of an option the user did NOT select is context, never a source — a value that exists only in unselected option text (a date boundary, a threshold, a gloss like "KST 기준") enters the spec as a recorded assumption or not at all. Context can carry facts nobody said (an earlier plan, a pasted doc, your own hypothesis options); the written ledger is what keeps them out of the spec.

## Requirement Clarification Summary

### Before (Original)
"{original requirement verbatim}"

### After (Clarified)
**Goal**: [precise description]
**Reason**: [why this matters]
**Scope**: [included and excluded]
**Constraints**: [limitations and preferences]
**Success Criteria**: [how completion is verified — push for a measurable threshold]

**Decisions Made** (user-selected):
| Question | Decision |
|----------|----------|

**Assumptions** (recommended defaults the user deferred — flag before building on them; omit if none):
| Topic | Assumed |
|-------|---------|

**Surfaced Along the Way**: [considerations the user hadn't raised, now decided or assumed — omit if none]

**Still Open**: [residual unknowns + how each gets resolved (experiment, measurement, later decision) — omit if none]

The user should leave with a better map, not just answered questions — the last two sections are where the interview pays for itself.

Write all content (activation message, questions, summary values) in the conversation's language; keep the structural headings and the promise tag verbatim so hooks and tooling can parse them.

Then output the terminal promise:

• <promise>CLARIFICATION COMPLETE</promise> — clarification is genuinely complete and the interview ended itself with round budget to spare.
• <promise>CLARIFICATION CAPPED</promise> — the round budget was fully consumed (whether or not material axes remain unasked — Phase 4 triage). Same terminal effect, honest label; Still Open carries the un-asked axes, or the single budget-exhausted line. The Stop hook enforces this: a COMPLETE emitted at the cap is rejected once with a request to re-emit as CAPPED.

After emitting either tag, leave .claude/clarify-interview.local.md and .claude/clarify-interview.guard in place — the plugin's Stop hook validates the tag against the loop state (a fully-consumed budget must not end COMPLETE) and deletes both files itself; deleting them in the same turn would blind that check. Only in a hookless runtime (no Stop hook installed — e.g., this SKILL.md executed standalone by a harness that says it owns the lifecycle) do you delete them after the promise. Cancellation deletes both files directly in both runtimes.

Loop Mechanics

AskUserQuestion returns answers mid-turn, so all rounds normally happen inside one turn and end with the summary + promise. The Stop hook is the safety net, not the driver: if a turn ends without a terminal promise tag, it re-injects the loop (re-read ## Clarification Progress first, then ask only what's still open) and enforces the 2h TTL; at max_iterations it demands the final summary once instead of asking further rounds; it deletes the state file when it sees either terminal tag. You own iteration while interviewing (step 4 above); the hook bumps it only on re-entry after an interrupted turn.

Now begin: create the state file, walk the territory, identify the material ambiguities and at least one blind-spot probe, and ask the first batch.