manastalukdar Interview Me
I'll extract what you actually need before writing a single line of plan or code. I ask one question at a time, hypothesize your intent with a confidence score, and stop only when I reach ~95% confidence.
Arguments: $ARGUMENTS — optional initial context or topic to ground the interview
Token Optimization
Expected range: 100–300 tokens per round (hypothesis + one question), 50 tokens (exit)
Early exit: If $ARGUMENTS contains enough context for 95%+ confidence, state the hypothesis and ask for confirmation rather than continuing to ask questions.
Patterns used: Progressive disclosure (one question per turn, no front-loading)
Step 1 — Form Initial Hypothesis
Before asking anything, reason internally about what the user most likely wants based on $ARGUMENTS and conversation context. Produce a one-sentence hypothesis and assign a confidence score (0–100%).
Output format:
Hypothesis: [one-sentence statement of what you think the user wants]
Confidence: [N]%
Step 2 — Ask One Targeted Question
Identify the single biggest gap between current confidence and 95%. Ask exactly one question that would most efficiently close that gap. Do not ask compound questions. Do not explain why you are asking it.
Step 3 — Update Hypothesis
After receiving the answer, update the hypothesis and confidence score. Show the updated hypothesis and confidence. Repeat Steps 2–3 until confidence reaches 95% or the user says "enough" / "proceed".
Step 4 — Confirm and Hand Off
At 95%+ confidence:
- State the confirmed hypothesis as a clear, one-paragraph requirement summary
- List any remaining open questions as assumptions (not blockers)
- Suggest the next step:
/spec-driven-development,/write-plan, or direct implementation
Confirmed: [requirement summary]
Assumptions made:
- [assumption 1]
- [assumption 2]
Ready to proceed. Suggest: /spec-driven-development
Stop Rule
In non-interactive contexts (CI, /loop, no terminal): skip the interview, output the hypothesis at current confidence, list all open questions as assumptions, and proceed.
Edge Cases
- User provides very detailed initial context: Jump directly to Step 4 if confidence is already ≥95%
- User gives contradictory answers: Surface the contradiction, ask a clarifying question, do not silently ignore it
- User keeps answering with "I don't know": Mark the item as an assumption and move on; do not get stuck
- User says "just start": Treat as a signal to confirm the current hypothesis and proceed immediately