Sololearn was looking for a new growth stream—the product taught programming and the team was testing whether it could expand beyond it. The EU AI Act would soon require companies to put employees through AI literacy training, which made teaching how to use AI the obvious direction. The idea was to reuse an existing format, adapt it, and test it.

The chosen format was Code Coaches: short practical exercises where learners wrote code, got hints from an AI mascot Kodie, and were checked against two criteria—syntax and whether they solved the problem.

Code Coaches worked because there was a right answer. Learning how to write a good prompt doesn't have one: there are clearer prompts and vaguer ones. The thing giving feedback was AI itself—autonomous, not supervised, capable of producing anything in the moment.

So the question wasn't only how to validate a prompt. It was how to teach someone to recognize a better one without pretending the system knew best.

The frame became: mistake-tolerant learning. The learner couldn't be wrong—they could be more or less specific. The product had to hold that distinction without flattening it into pass/fail and still make it feel like a lesson.

The first call we had to make was evaluation: to evaluate each prompt or the whole conversation. The team went with the second option to give learners play time: they sit in the prompting field, try things, iterate, and decide when they've practiced enough.

The second call was about feedback itself: to include or to skip. Learners were already prompting AI and if “another” AI provided feedback, two AIs had to be differentiated: the one they prompted and the one that coached them. The team had been ready to drop in-exercise hints entirely and only evaluate the response. But a learner getting no signal mid-exercise would leave with nothing to act on. The move was to place Kodie—already a known character to learners—above the prompt field as a small "Need a hint?" tap. Behind the scenes it was the same model. To the learner, it was a separate voice: the AI you prompt and the coach who watches you prompt it.

The third call was the language of the wrong state. The evaluation had two outcomes: the prompt fulfilled the task or it needed more details (vacation dates missing, tone too casual, intent unclear). Calling the second one "wrong" felt off. So the team used ‘Almost there. Let's see how to make your prompt better’ to indicate inaccuracy and invite iteration without black-and-white judgment.

Either way, the learner always had a choice: to complete the assignment without reaching a better answer or to go back and try again.

Three pressures pulled against each other. Shipping in two weeks. Reusing the existing system as much as possible: same components, same patterns, same mascot. And honoring the fact that the underlying logic of prompting was nothing like the underlying logic of code.

The resolution wasn't a compromise. It was using the familiar shell to carry a different mechanic: layout from Code Coaches, Kodie from Code Coaches, and evaluation flow shaped around the new constraints.

The AI Playground shipped. The mistake-tolerant frame held up against the unpredictability of the model—learners could explore without being penalized. The format was strong enough to spin into a new B2B product Sololearn later built for EU compliance training.

It was counterintuitive to let learners leave without reaching “the right” answer. But with a medium that's autonomous, unpredictable, and has no single correct output, the learning was never in the final score—it was in the iteration.