Meet the Speaker
Walk into a visit knowing exactly who to meet and what to say.
You are about to spend a week among people whose papers you cite — let’s make sure you actually talk to the right ones, about the right things.
The goal
The Goal Your lab has been invited to give a talk at a renowned pain-research university (say, McGill). By the end of this challenge you should be able to walk onto that campus with two things in hand: (1) a proposed talk title plus a ranked shortlist of faculty to meet — each with a one-line reason rooted in genuine research overlap — and (2) a “match scorecard” for the speakers and hosts on your itinerary, so every hallway conversation has a real opener instead of “so… what do you work on?” How you get there is up to you.
Why it matters
The science you remember from a conference is rarely the keynote. It is the 6-minute conversation by the coffee urn where someone said “you should look at our unpublished pilot” and your next grant aim was born. But those conversations do not happen by accident — they happen when you walked in knowing that this person just pivoted from rodent models of neuropathic pain to human QST, and that your dissertation has something to say about it.
Doing this homework by hand is slow: dozens of faculty pages, Google Scholar profiles, lab sites that were last updated in 2019. A research-capable AI can compress hours of roster-scraping into minutes — drafting overlap maps, candidate talk titles, and conversation openers. The catch, and the whole point of today: AI confidently invents affiliations, fabricates papers, and mixes up two researchers with the same surname. Your job is to move fast and then verify carefully. That is the same habit that keeps you out of trouble in a literature review or a grant background section too.
Run of show
- 0:00–0:05 · Challenge introduction (5 min)
- 0:05–0:20 · Work in your group (15 min)
- 0:20–0:22 · Post your best prompt (2 min)
- 0:22–0:32 · Share & debrief (10 min)
- 0:32–0:35 · Reset (3 min)
Bad prompt to better prompt
Why it disappoints: you get a generic, possibly outdated list of famous names (and at least one who moved institutions or retired), zero connection to your work, no ranking, no reasons, and nothing you could actually open a conversation with. It reads like a generic Wikipedia summary, and you have no way to tell which parts are hallucinated.
Why it works: it gives the AI an identity and your research fingerprint (so overlap is real, not generic), demands a structured + ranked output you can act on, forces self-scored confidence and an explicit verification list, and pre-empts fabrication. You leave with an actionable plan and a fact-check checklist in one pass.
Prompting moves to try
- Decompose the goal. Split it: (1) list candidate units, (2) list faculty per unit, (3) score overlap, (4) draft openers, (5) draft a talk title. Run them as separate turns so each step is checkable instead of one mega-answer you can’t audit.
- Role / identity prompting. “You are a research-strategy advisor” or “you are a skeptical senior PI” changes the output more than you’d expect. Also feed it your identity — your methods, model system, and hot question — so “overlap” means something.
- Adversarial self-evaluation. Ask the AI to attach a 0–100 confidence score to every affiliation and paper, then to list its three least-trustworthy claims. Make it argue against its own shortlist: “Which of these matches is weakest and why?”
- Ask the AI to improve your prompt. “Here is my prompt. What’s missing that would make your answer more accurate and more useful to me? Rewrite it, then answer the improved version.”
- Make it cite or it doesn’t count. Require a source/link per faculty claim. No retrievable source = treat as a hypothesis to verify, not a fact.
- Generate the opener, not just the bio. Push for a question only an insider would ask (“Has your group looked at whether the QST effect survives covarying for catastrophizing?”) rather than “I enjoyed your work.”
Starter materials
Pick a track or do both. Everything below is paste-ready.
Track A — The talk visit (faculty match)
Paste this scenario and roster snippet into your AI and have it rank, score, and propose a talk title. (Names below are fictional composites built for this exercise — treat the AI’s “corrections” to them with extra suspicion.)
Your lab, in one paragraph We are a translational pain lab. Core program: sex differences in descending pain modulation in chronic low-back pain, combining 3T fMRI (periaqueductal gray / rostral ventromedial medulla connectivity) with quantitative sensory testing and conditioned pain modulation. Side interest: how expectation and catastrophizing shape placebo analgesia. We’re presenting next month and have ~5 slots for one-on-one meetings.
| Faculty (fictional) | Unit | Stated focus | Recent direction |
|---|---|---|---|
| Dr. A. Moreau | Alan Edwards Centre for Research on Pain | Spinal mechanisms of neuropathic pain, rodent | Moving into human biomarker validation |
| Dr. R. Okafor | Psychology | Pain catastrophizing & expectation | Placebo analgesia in clinical trials |
| Dr. L. Tremblay | Anesthesia | Opioid-sparing perioperative protocols | Central sensitization imaging |
| Dr. S. Nakamura | Neurology | Migraine & cortical excitability | Sex hormones × pain thresholds |
| Dr. P. Singh | Biomedical Engineering | MR methods, brainstem imaging | High-res PAG/RVM connectivity at 7T |
| Dr. C. Beaulieu | Nursing | Self-report measures, chronic LBP cohorts | Large longitudinal LBP dataset |
Deliverable A A ranked shortlist (top 5) with a meet-priority score and reason for each, ONE proposed talk title that lands with this audience, and a flagged “verify-before-you-trust-me” list (affiliations, any paper the AI names, who actually still works there).
Track B — Party Planner (rapid speaker recon)
You have an itinerary. Have the AI fill a scorecard so you can talk to each person like you’ve read their last three papers — because functionally, you have.
Sample itinerary Tue 10:00 — Keynote: Dr. Okafor, “Expectation as an Analgesic” · Tue 13:30 — Lunch table host: Dr. Singh · Wed 09:00 — Panel: Drs. Moreau & Tremblay (translation: bench to bedside) · Wed 16:00 — Poster session walk-through with Dr. Beaulieu · Thu 11:00 — Career-chat slot: Dr. Nakamura (early-career awardee).
Match scorecard template — ask the AI to return one row per person:
| Person | Their big question (1 line) | Overlap with me | Conversation opener (specific) | Avoid / sensitive | Confidence 0–100 | Must-verify |
|---|---|---|---|---|---|---|
| e.g. Dr. Okafor | Does expectation reduce pain via descending modulation? | I image the PAG/RVM circuit she infers behaviorally | “Have you tested whether the expectation effect tracks PAG-RVM connectivity?” | Don’t pitch as collaborator first contact | 72 | Confirm she gave THIS keynote; confirm placebo paper exists |
Deliverable B A filled scorecard for all five itinerary people, plus your single highest-confidence opener for each — and a flag on any row scoring under ~60 that you would NOT say out loud until you’ve checked it.
Verify, then trust Research-capable AI will state a wrong current affiliation with total confidence, cite a paper that does not exist, and confuse two researchers who share a surname. Treat every name, title, and citation as a lead, not a fact. Before you walk into a room: confirm the person still works there and actually appears on your itinerary, and never quote a “paper” the AI named until you’ve found it yourself. Getting someone’s work wrong to their face is worse than not knowing it.
Debrief questions
- Which AI’s shortlist had the strongest reasons — and did the highest-ranked person survive a fact-check, or evaporate?
- What did the AI get confidently, specifically wrong? (Affiliation? A fabricated paper? A merged identity?) How would you have caught it if you hadn’t been looking?
- Did role-prompting (“skeptical senior PI”) or feeding your own research fingerprint change the overlap quality more?
- Which conversation opener would you actually say out loud — and which one would embarrass you if the underlying fact were wrong?
- How much of this could you reuse for the background section of a grant or a literature review, and what’s the same verification burden?
Level up
- Two-pass adversarial recon. Generate the shortlist with one AI, then paste it into a second and ask it to find every error, weak match, and unsupported claim. Compare what each model is willing to admit it doesn’t know.
- Talk title A/B/C. Have the AI produce three talk titles tuned to three different audiences in the same building (basic neuroscientists, clinicians, methods/imaging people) and justify which slot each fits.
- Build a reusable conference kit. Turn your best prompt into a saved template with slots for [my work], [target institution], [itinerary] so you can re-run it for every meeting this year — and bolt a standing “confidence + must-verify” section onto it permanently.