The writing desk wasn’t there.
It was the fall of 2008, and I was playing Valmont in a college production of Dangerous Liaisons—my first time in heels and a powdered wig. My heart sank as my scene partner and I realized that this desk, critical to our blocking, was nowhere to be seen. A stagehand had missed their cue. The desk was still in the wings behind a trellis on squeaky casters. The scene was already in motion. We had about two seconds to decide: break, or bend.
We bent. The actress playing Merteuil reclined onto the chaise, beckoned me to sit beside her, and we proceeded with the scene mere inches away from one another. Our lines crackled with a new energy and intimacy. Audience members talked about it afterward. Two weeks into our four-week run, we kept the new blocking.
I think about that desk a lot when I work with AI now.
Find your collaborators
I spent more than two decades in the theatre before I came to L&D. Most of that time I was around people whose job is, very practically, to make something out of nothing—actors, designers, directors, dramaturgs, stage managers, the techies who hang lights at three in the morning. One thing you learn quickly in that world is that theatre is not an autocracy. The director has a vision, but the play that ends up on stage is always —always— a negotiation between the people in the room.
The first design meeting on a new production is a small revelation every time. You walk in with one picture of the play in your head, and you walk out with another, because the lighting designer has seen something you didn’t, and the composer has heard something you didn’t, and the costume designer has built a world you didn’t know was possible inside the world you thought you were making. You get out-imagined, in the best possible way. The play gets bigger than you.
In grad school, Oskar Eustis, the Artistic Director of the storied Public Theater, gave my cohort what he called his best piece of advice to aspiring dramatists: find your collaborators. He meant the people whose visions align with yours, and the people whose visions productively oppose yours. The ones who push you off the path you would have walked alone, and onto a more interesting one.
Generative AI, used well, is exactly that kind of collaborator.
The dication trap
A lot of how people talk about working with AI right now sounds like dictation. You write a prompt, the model “delivers” what you asked for, and you grade the output against the picture in your head. If it doesn’t match, you tighten the prompt and try again. The whole interaction is shaped by the assumption that your original vision is the target and the AI’s job is to hit it.
This is how an unskilled director works with an ensemble. And from firsthand experience, it rarely leads to a great outcome, neither for the creatives nor the audience.
When I prompt a model now—to draft scenario branches for a learning module, to riff on linguistic directions for a character, to propose mechanic variants for a minigame—I try to come in with a clear intention and a loose vision. I know what the moment needs to do. I’m less precious about exactly how it does it. Then I read what comes back the way I’d watch a scene partner make an unexpected choice on stage. Not “is this what I asked for,” but “what is this giving me that I didn’t see before?”
Sometimes the answer is nothing, and you cut it. But often there’s a turn in there —a phrasing, a structure, an angle— that I would not have arrived at on my own in a week of trying. And the smart move, the move I learned on a university stage in 2008, is to take what’s being offered and build with it.
Why this matters for learning design
Because surprise is one of the most reliable tools we have for making something stick.
Cognitive scientists have studied this for decades under several different names—prediction error, expectation violation, the novelty effect. The short version is well established: when what happens next doesn’t match what your brain predicted would happen, attention sharpens and the brain flags the moment as worth encoding. Novelty recruits the same dopaminergic systems involved in memory consolidation. Distinctive moments inside an otherwise predictable sequence are recalled disproportionately well; the phenomenon has been replicated so many times it has its own name (the von Restorff effect, if you want to look it up).
The training you remember is the training that surprised you
If you’ve ever sat through a workplace training course, you already know the practical implications. The material that taught you nothing was the material you could predict. The moments you remember are the ones that didn’t go the way you expected.
This is the through-line, for me, between theater and L&D. The job is not to deliver content. The job is to engineer a moment that lands. And moments that land almost always involve some small subversion of what the learner thought was about to happen next.
If I am the only person in the room when I’m designing those moments —just me and my own pattern-matching brain— I will produce variations on what I’ve already produced. The unexpected is, by definition, hard to generate from inside your own expectations. This is precisely what a good collaborator gives you. And it is what AI, used as a collaborator and not a vending machine, gives you several times an hour.
Don’t let perfect be the enemy of good
When I approach AI with a collaborative spirit, I tend to adhere to a simple framework drawn from how a skilled director works with an ensemble:
Bring an intention, not a mandate. Know what the moment needs to do for the learner. Hold the how loosely. The tighter your prompt prescribes the output, the more you constrain the model to produce variations of what you already imagined.
Treat every response as an offer. In improv, we say “yes, and...” The first move is always to receive what the scene partner is giving you on its own terms before you decide what to do with it. When AI returns a draft, an image, a structure—read it for what it suggests, not just for whether it matches the picture in your head. Remember: an AI’s output is in response to something you gave it, and it’s possible that an unexpected return could indicate blind spots, biases, and assumptions you didn’t know you had
Direct the whole. You’re not the writer anymore; you’re the creative director. Your taste, your judgment, your read on the learner and the brand and the moment—those are still entirely yours, and they’re what the work actually depends on. The AI proposes. You shape, cut, combine, and decide.
That’s it. The framework is short on purpose. Like most things in theater, it works in practice and gets pretentious when over-explained.
There’s an obvious objection to all of this: that being open to the unexpected sounds incompatible with shipping fast. It isn’t. At Attensi, we move quickly and iterate, and one of the reasons we can is precisely because the designers don’t try to white-knuckle every output into matching a fixed vision. We let the tools surprise us, we keep what’s strong, we cut what isn’t, and we move. “Don’t let perfect be the enemy of good,” is theatre advice older than I am. So is the rule that eighty percent of something live beats a hundred percent of something that never opens.
The desk wasn’t there. We made a better scene.
An interesting and sensitive question for L&D right now is not whether AI will replace designers. It won’t, for the same reason an excellent ensemble doesn’t replace its director. The more interesting question is whether we’ll stay open enough to let AI actually collaborate with us—whether we’ll keep our visions firm enough to lead and loose enough to be inspired.
Find your collaborators. Some of them, now, are large language models.
Curious how other creatives in L&D are framing this. Is your AI a vending machine or a scene partner?
