| name | interrogation |
| description | Use when the user invokes `/interrogation "<prompt>"`. Ask focused clarifying questions via AskUserQuestion until you have enough certainty to act on the prompt without guessing. Stop early — three rounds at most. |
Interrogation
The user gave you a prompt that's likely under-specified. Drill in with structured questions before doing the work, so the output lands on the first try instead of needing a redo.
Workflow
- Read the prompt carefully. Identify the 2–4 ambiguities that, if guessed wrong, would lead to materially different output. These are the only things worth asking about.
- Ask via
AskUserQuestion. Batch related questions into one call — never one question per turn. Use multiple-choice when the answer space is bounded; use a free-form option for open ends.
- Outline your understanding after every answer. Don't silently decide "I have enough." Write a short, high-level outline of what you now think the user wants — scope, output shape, key choices — so the user can see your mental model and correct it before more time is invested. Bullet points are fine. Keep it to ~5 lines.
- Iterate at most twice more. If the outline reveals a remaining ambiguity, ask another batched round. If the user confirms the outline (or says "go" / "ship it"), stop. Three rounds maximum.
- Final confirm and proceed. Once the outline matches, restate the agreed plan in one paragraph and begin the work.
What to ask about
- Scope — what's in, what's out
- Output shape — file? PR? branch? message? what does "done" look like?
- Hard constraints — must use X, can't touch Y, deadline, audience
- Judgement calls the user has a view on — speed vs. polish, convention vs. novelty
What NOT to ask
- Anything answerable from the codebase, recent context, or
CLAUDE.md
- Stylistic micro-decisions — pick the codebase convention silently
- Things the user already answered earlier in the session
- Yes/no when a multiple-choice would be more decisive
Stop conditions
- The user confirms your outline (mental models match)
- The user says "go", "ship it", "you have enough", or equivalent
- Three rounds in, further questions would scope-creep rather than clarify
