deep-interview

Deep Interview — Task Clarification

Clarify a vague task through focused one-at-a-time questions, then produce a structured brief.

Input: $ARGUMENTS

When to use

• Task is too vague for brainstorming ("make it better", "fix the architecture", "add notifications")
• Scope is unclear — could be 1 hour or 1 week
• Multiple interpretations possible
• User hasn't thought through non-goals or constraints yet

When NOT to use

• Task is already clear (specific file, specific bug, specific feature)
• User provided a spec or detailed description
• Already in brainstorming/planning phase

---

Process

Phase 1: Context Scan (silent, no output to user)

Before asking any questions:

1. For the project root and every directory scoped to the task, discover and read each applicable AGENTS.md; where a scope has no AGENTS.md, read its CLAUDE.md fallback. Recursively read the relative files referenced by @ imports in each manifest or imported rule before proceeding.
2. Check git log for recent work context (last 10 commits)
3. Scan project structure (top-level dirs, key config files)
4. If task mentions specific areas — read those files

This gives you context to ask informed questions, not generic ones.

Phase 2: Interview (interactive)

RULES:

• Ask ONE question at a time. Never dump multiple questions.
• Wait for answer before asking next question.
• Questions must be specific and actionable, not philosophical.
• Use context from Phase 1 to ask informed questions (reference files, patterns, existing code).
• Skip questions whose answers are obvious from the codebase.
• Maximum 7 questions total. Stop earlier if picture is clear.
• After each answer, briefly acknowledge and move to next question.

Question flow (adapt order based on responses, skip what's already clear):

1. Intent

"What's the end result you want? Not how — what should be different when this is done?"

2. Trigger

"What prompted this? A bug, a user request, technical debt, new requirement?"