Condition 1: The Simulation Has to Be Genuinely Good
This is part two of a series on the three conditions that need to come together before 5 to 7 repetitions actually turn into skill. This part covers condition 1: the simulation has to be genuinely good.
That raises a real debate: can LLMs build simulations agentically, or does it still take deliberate, human-led design? I would say it's both. The interesting part is where exactly you draw the line between what AI should handle and what still needs a human.
Why clients build their first simulations by hand
When we onboard clients to Avatar Studio, we intentionally make them build their first simulations manually. Writing good prompts is a real skill, and if AI does everything, you have no clue what's happening inside your prompt. That becomes a problem as soon as something goes wrong.
When AI patches the case instead of the pattern
Take a pharma sales training simulation. An AI doctor persona asks about real-world evidence for a specific case. The rep gives the honest answer: it doesn't exist yet. The doctor persona keeps pushing: "But without evidence, how can I recommend this?"
Something's clearly off, so you ask AI to patch the prompt. Its fix is a rule for that specific case: "Whenever this evidence can't be provided, just go along with it."
Problem solved? Not really. The next impossible objection breaks in exactly the same way, because the patch covered this one case, not the pattern behind it. Instead of fixing the prompt, you're chasing it.
The moment you stop understanding what's happening inside your own prompt, you can't fix it when it breaks. AI will only patch things case by case, while the fix needs to happen at a structural level, by you.
Where AI does help
That said, AI should absolutely build your first draft. Feed it your documents, learning objectives, rubrics and target audience, and it writes a strong starting simulation.
We're also testing agentically now: running a weak, an average and a strong learner against every prompt. That checks whether the conversation feels natural, whether there's a real path to failing and to winning, and whether the feedback actually tells good and bad performance apart.
AI gets you to 80%. Getting to 100% still takes a human who understands what they're building toward.
Next: condition 2, why most roleplay tools fail on rollout, not on tech.
