There’s a specific kind of meeting where you walk in underprepared — not on the subject matter, but on the people. You know the topic cold. You don’t know whether the person across from you is going to be defensive about budget, skeptical about the timeline, or quietly annoyed about something that happened two months ago that nobody officially discussed. You find out mid-sentence, when it’s too late to adjust.
AI is surprisingly useful for this. Not because it knows the people involved — it doesn’t — but because it’s good at generating the space of likely dynamics once you give it enough context about the situation.
What You’re Actually Asking It to Do
The move isn’t “tell me how this meeting will go.” That’s too vague and the output will be useless. The move is closer to: give me a cast of characters, each with a plausible read of this situation, and tell me what each one probably wants and is worried about.
Concretely: you’re about to pitch a project change to a senior stakeholder who approved the original scope. You tell Claude that. You describe what the change is, roughly what it costs, and what the original reasoning was. Then you ask it to play out the stakeholder’s likely concerns — not as a simulation, but as a structured list of what someone in that seat usually needs to hear, and what usually triggers pushback.
The output isn’t magic. But it surfaces three or four objections you weren’t foregrounding, and at least one of them is probably real. That’s the value — not prediction, but coverage. You’re using AI to audit your own blind spots about a social situation the same way you’d use it to audit your own blind spots about an argument you wrote.
The Prompt Pattern That Actually Works
Vague prompts get you vague dynamics. “How might this meeting go?” produces a hallucinated neutral version of every meeting ever. What works is giving the model a specific seat to inhabit and a specific constraint to reason against.
Something like this:
“I’m meeting with [role, not name] to propose [specific change]. Their last interaction with this project was [brief context]. They’ve historically prioritized [X]. What are the three most likely objections they raise, what’s the underlying concern behind each one, and what would actually address that concern — not just deflect it?”
The “underlying concern” piece is the part most people skip. It forces the model past surface objections (“the cost is too high”) into the actual worry underneath (“if this goes over budget it reflects on my judgment for approving it”). Those are different problems with different solutions.
You can run this for group dynamics too. If there are two people in the room with different priorities, ask the model to generate each person’s likely read of the same proposal, side by side. Not to predict an argument, but to spot where you’re going to need to phrase something twice.
Where It Actually Falls Apart
It doesn’t know the actual person. This sounds obvious but it’s easy to forget once you’re reading a confident-sounding list of “likely concerns.” What the model is generating is the statistical shape of someone in that role, not the specific human. A CFO who is personally excited about the project direction is not well-modeled by “CFOs are typically focused on cost containment.”
So the output has to be treated as a checklist of possibilities, not a prediction. You’re looking for coverage — did I think about the budget angle? did I consider how this lands for someone who wasn’t in the original scoping conversation? — not a forecast. The moment you treat it as “here’s how it will go” instead of “here are angles I might have missed,” you’re using it wrong.
There’s also a subtler failure mode: AI-generated stakeholder prep tends to be rational. It surfaces logical concerns, budget concerns, timeline concerns. It’s worse at surfacing the irrational-but-real ones — the person who’s annoyed because they feel excluded from a decision, or the history between two people in the room that you don’t know about. Those are the things that actually derail meetings, and they don’t show up cleanly in a structured prompt response. You have to hold space for those manually.
A Few Specific Situations Where This Works Well
Salary conversations. If you describe the situation — your role, what you’re asking for, roughly what the company’s constraints seem to be — the model will generate a reasonably useful map of where pushback is likely to land and what framing tends to work against each type of objection. It’s essentially doing what a mentor with industry experience would do, except it’s available at 11pm the night before.
Client calls where you’re delivering bad news. These have a fairly consistent shape — client hears the problem, moves through a predictable emotional arc, asks a predictable set of questions — and AI is reasonably good at laying out that arc so you can figure out where to get ahead of it rather than just react.
Pitches to people you don’t know well. When you have limited signal about someone and you’re about to make an ask, generating a few plausible versions of this person based on their role and context is genuinely useful for deciding how to open and what to lead with.
How to Calibrate Your Trust in the Output
One useful test: after the model gives you a list of likely concerns, ask it which one it’s least confident about and why. If it identifies where it’s generalizing from role patterns rather than from actual signals you gave it, that’s a sign you can lean on the rest. If it just repeats the list with different words, treat the whole output with more skepticism.
Another: run the prompt on a meeting that’s already happened and check whether the concerns it surfaced match what actually came up. If it’s consistently missing the type of thing that matters in your context, that’s useful calibration for how much weight to put on it going forward.
Honestly: it’s better than going in with nothing, and worse than a real conversation with someone who actually knows the room. Most of the time you have neither, so better than nothing is enough.
Before your next high-stakes conversation, try writing out what you think the other person wants and is worried about before you prompt anything — then compare that list to what the model generates. The delta between the two is almost always more interesting than either list on its own.
