ramsbaby/jarvis/deep-interview

/deep-interview — Convergent Requirements Interview

Acts as a senior architect. 8-15 rounds of Socratic questions narrow a vague idea into a production-grade spec. Blocks contradictions, edge-case gaps, and security gaps before any code is written.

How it differs from /brainstorm

• /brainstorm (divergent): expand possibilities. 15-20 questions + UI mockup variants. Exits when ideation feels saturated.
• /deep-interview (convergent): spec gating. 8-15 narrowing rounds + decision log. Exits when ambiguity score ≤ threshold.

When to use which:

• Idea itself is undefined → /brainstorm
• Idea exists, implementation decisions are undefined → /deep-interview

Mechanism (4 stages)

1. Mathematical ambiguity gating

After each round, score across:

• Undefined behavior areas
• Contradiction risk (conflicting requirements)
• Uncovered edge cases
• Security gaps (secrets, PII, permissions)
• Unidentified irreversible decisions

Refuse autonomous implementation until score ≤ 0.2 (default threshold).

2. Socratic question patterns

Examples the agent will ask:

• "What happens when X scenario triggers Y?"
• "What evidence would prove this decision wrong (falsifiability condition)?"
• "What assumptions are you carrying that aren't yet stated?"
• "What technical debt will this create one year out?"

3. Contrarian / Simplifier Challenge Mode

For each major decision, automatically run two checks:

• Contrarian: deliberately raise the counter-hypothesis (debiasing — never confirm a single hypothesis without testing its opposite)
• Simplifier: check whether a much smaller version exists (Karpathy "Simplicity First" — 200 lines vs 50 lines)

4. Decision log auto-write

All decisions, rationale, falsifiability conditions, and challenge results are recorded in JSONL for later audit.

Iron rules integration

• No fix without root cause: 5-why on each decision
• Never lie about status: "don't know" is explicitly marked, not papered over
• User sovereignty: agent questions are strong signals, not commands — user's answer is final

Self-check (per round — BLOCKING)

1. Did I compute the ambiguity score with concrete evidence?
2. Did I avoid confirming a single hypothesis without contrarian check?
3. Did I run the simplifier challenge?
4. Did I write the decision and its rationale to the log?

All must pass to advance to the next round.

Limitations

• Too-divergent input → recommend /brainstorm first
• Rounds > 15 → auto-pause with "convergence failed, recommend re-divergence"
• Ambiguity score is a heuristic, not a precise measurement — user judgment is final

Related skills

• /brainstorm — divergent ideation (predecessor stage)
• /plan-review — rigorous spec review after this skill completes
• /verify — implementation verification after spec is built
• /office-hours — trade-off matrix when a single decision needs explicit user approval

Source attribution

Inspired by Sorbh/interview-me (github.com/Sorbh/interview-me). Convergent gating concept and Contrarian Mode adapted with permission from the open-source pattern.