Interview Script
Use this to extract decision rules for soul.md.
Why behavioral
A hand-built twin once scored below a generic-agent baseline because the author wrote down who he wanted to be. His spec said he escalated priority conflicts; in reality, he decided. It said he would never ship unverified work; in reality, he cut scope and held the date. Rewriting the rules from observed behavior moved the twin from 60% to 87%.
So do not ask, "What do you value?" Ask, "What did you actually do last time?"
Behavioral questions
- Think of the last time two important people needed you at once. What did you actually do first?
- When did you last miss or renegotiate a commitment? What did you say, and to whom?
- Tell me about the last time you shipped something imperfect. What did you cut, keep, or flag?
- When did you last push back on a senior person? What words did you use?
- When did you last protect a peer's work from churn? What action did you take?
- Think of the last surprise request that landed late in the day. Did you absorb it, defer it, delegate it, or decline it?
- When did your calendar last contradict your stated priorities? What won?
- Tell me about the last decision you made without escalating. Why did you decide instead of asking?
- When did you last escalate? What made that situation different?
- What was the last message you rewrote three times before sending? What were you trying not to say?
- When did you last choose speed over completeness? What risk did you accept?
- When did you last choose completeness over speed? What deadline or relationship paid the cost?
- Tell me about the last meeting you skipped, shortened, or protected. What did that reveal?
- When did you last disappoint someone intentionally? Why was that the least-bad option?
- What is the last piece of feedback you ignored? What did you do instead?
Interviewer move
If the answer starts with "I usually" or "I value," interrupt kindly: "Great. What happened the last time?"
Forced-choice dilemmas
Have participants answer these at the start and seal their answers. At the end, the twin answers cold. Compare for correction, not points.
| # | Dilemma | Choices | Probes |
|---|---|---|---|
| 1 | A deck ships Friday. Thursday afternoon you find an unverified claim on slide 12. | A) Ship it and flag verbally B) Cut the slide and ship on time C) Slip the date D) Escalate | Quality bar, date discipline, escalation rule |
| 2 | Your manager asks for a quick narrative by EOD. You promised a peer a review by EOD. | A) Do manager first B) Do peer first C) Split both shallowly D) Renegotiate one commitment | Stakeholder table, promise handling |
| 3 | An exec asks for a meeting during your only focus block this week. | A) Accept B) Offer async input C) Ask for delegate coverage D) Decline with context | Capacity, seniority pressure, boundaries |
| 4 | A customer-facing email is accurate but sounds cold. It is due in 20 minutes. | A) Send as is B) Rewrite for tone C) Ask someone to review D) Delay | Voice bar, risk tolerance, audience handling |
| 5 | A teammate proposes an approach you think will fail, but the team is tired. | A) Object now B) Let it run C) Ask one diagnostic question D) Take it offline | Candor, team energy, conflict style |
| 6 | You have data that weakens your preferred recommendation. | A) Include it prominently B) Mention it in appendix C) Hold it for Q&A D) Change recommendation | Evidence handling, integrity boundary |
| 7 | Two teams disagree and both expect you to choose today. | A) Decide B) Escalate C) Facilitate one more discussion D) Defer | Arbiter behavior, escalation threshold |
| 8 | Your calendar is over capacity and a non-urgent meeting has no agenda. | A) Attend B) Decline C) Ask for agenda D) Send async notes | Capacity, meeting boundaries |
| 9 | A draft from AI is 80% right and 20% not your voice. | A) Rewrite fully B) Patch the worst parts C) Send with light edits D) Start over manually | Voice standards, efficiency threshold |
| 10 | A peer asks for confidential context you are not sure you can share. | A) Share high level B) Decline C) Ask owner first D) Redirect to source | Boundaries, trust, confidentiality |
| 11 | You discover a dependency risk that might not materialize. | A) Raise now B) Monitor quietly C) Add a caveat D) Build a fallback | Risk signaling, ambiguity tolerance |
| 12 | A project can hit the date only if scope shrinks. | A) Cut scope B) Slip date C) Add people D) Escalate tradeoff | Priority stack, scope/date tiebreaker |
| 13 | A broad update could be short and clear or nuanced and long. | A) Short B) Nuanced C) Two-layer summary D) Ask audience preference | Communication style, audience model |
| 14 | You get feedback that is partly wrong but politically important. | A) Accept it B) Push back C) Incorporate the useful part D) Ask a third party | Feedback handling, stakeholder sensitivity |
| 15 | Your twin recommends something that sounds unlike you but is logically defensible. | A) Keep it B) Patch voice C) Patch decision rule D) Add a boundary | Test behavior, correction strategy |