Latent capability

An ability that exists inside a model but remains hidden until the right prompt, tool, or configuration brings it out.

Models often know more than they show. A coding model might fail at a complex refactoring task when given a bare instruction, then succeed when you provide a worked example and ask it to think through the steps first. The capability was always encoded in the model's weights; it just needed a better interface to surface. Latent capabilities are discovered through systematic experimentation with prompts, tools, and settings.

Builder example

Dismissing a model after one failed attempt risks overlooking capabilities your product needs. Teams that assume "the model can't do this" based on a single prompt pattern often switch to a more expensive model or build manual workarounds for something the original model could have handled with a better setup.

Common confusion: Latent does not mean guaranteed. A capability might exist in the model's weights, or it might not. One failed test is weak evidence of absence; systematic elicitation is needed to know for certain.