Interview prep from primary source.
The three questions worth asking, drawn from what they've actually written or said — not their Wikipedia page.
You have a 30-minute interview with someone whose work you should know but haven't read deeply. Paste their published work — a paper, a blog series, a podcast transcript — and this prompt returns the specific questions that show you actually read them.
THE PROMPT
I'm interviewing [NAME] for [OUTLET/PURPOSE]. Attached is a body of their published work — [PAPERS/POSTS/TRANSCRIPTS]. Prepare me by returning: **Their argument, in one paragraph** The core position they hold across this body of work. Not what they're "known for" — what they've actually argued. If they've changed their mind over time, note that. **Three questions worth asking** - One question that engages with the strongest version of their argument and pushes on it. - One question that surfaces a tension between two of their positions across the attached work. - One question that would move the conversation forward — asking about something they haven't yet published on but which follows from their framework. For each question: state the question, then in one line, note which specific text/quote you're drawing from. If the question requires them to update or defend a position, name that. **One question NOT to ask** The generic question that shows you haven't read them. Something they'd have answered 100 times. Save yourself the reputational hit. **Warm-up angle** One sentence — the low-stakes opening question that gets them talking without being a softball. Attached: [PASTE PAPERS OR TRANSCRIPTS]
**Their argument**
Across their four papers and the two podcast interviews, [Name] argues that the standard framing of "AI safety" as adversarial (misaligned models vs. controllers) misses the more common failure mode: systems that are subtly misaligned in ways that compound over deployment cycles without triggering safety trips. They've refined this from a purely technical framing (2023 paper) to one that includes org-design factors (2025 essay).
**Three questions worth asking**
- In the 2025 essay you argue the compounding-misalignment framing implies specific org-design changes at labs. Which of those has actually been adopted at any of the frontier labs, and what does the resistance tell you? — draws from Section 4 of "Compounding Misalignment".
- In your 2023 paper you argue evaluation benchmarks are the wrong instrument for detecting subtle misalignment; in your April 2026 podcast you praised the new [X] benchmark suite. How do you reconcile those? — a real tension worth surfacing.
- If the compounding-misalignment framing is right, what should a regulator specifically require of deployers (not providers)? You've been quiet on the deployer side of the value chain.
**One question NOT to ask**
"How worried are you about AGI risk on a scale of 1 to 10?" — they've fielded this dozens of times and it teaches your audience nothing new.
**Warm-up angle**
Ask them how their frame shifted between the 2023 paper and the 2025 essay — an easy opener that gets them describing their own intellectual history, which they'll enjoy, and which sets up the harder questions.