ColdStart: the hardest moment with an AI is the first one
Most people leave the majority of a model’s capability unused. Not because the model cannot do it, but because they do not know how to ask. That gap is what ColdStart is being built to close.
A football scout knows exactly which player profile they need. A shop owner knows exactly what is going wrong with their margins. A student knows exactly what they do not understand. All three have real expertise and a clear goal. What none of them has is the vocabulary to translate that into something a model can work with efficiently.
The result is a gap that is widening rather than closing: models keep getting more capable, and the average first prompt does not. Researchers call it the First Prompt Problem, and the consistent finding across pilot groups is that the hesitation is not about capability. People understand what the model can do. They stall at how to begin.
The industry’s answer so far has been documentation, tutorials, prompt libraries and better tooltips — all of which put the burden back on the person who already told you they do not know where to start. ColdStart takes the opposite route: instead of teaching the user to ask better, it removes the requirement to know how.
The part we find most interesting is the classification. ColdStart does not sort people by profession, because profession predicts very little about the friction someone is actually experiencing. It sorts by the nature of the gap — the person who knows exactly what they want but not how to say it is a completely different problem from the person whose goal is not yet clear even to themselves. Those two need opposite treatments, and most systems give them the same blank box.
There is an obvious question attached to any system that builds a profile of the person using it: where does that profile live. It is a fair question and we would rather answer it before anyone has to ask. The observation layer runs on local embeddings, and the configuration step can run entirely on local inference. A profiling system built by a company that sells data sovereignty would be a poor advertisement for it otherwise.
Where it stands: phase zero, architecture and definition. There is no public repository and nothing to download. We are writing about it now because the problem it addresses is real and badly served, not because the answer is finished.