Capability elicitation

Finding the right prompt, tool setup, or configuration to unlock abilities a model already has but does not show by default.

Capability elicitation is the systematic process of finding the right prompt, tool setup, or configuration to unlock abilities a model already has but does not show by default. Techniques include providing worked examples, giving tool access, breaking the task into steps, or adjusting temperature and other settings. Early large language models seemed unable to do multi-step math until researchers discovered that adding "let's think step by step" to the prompt unlocked the ability. The capability was always there; it needed the right trigger.

Builder example

If you test a model with one prompt style and conclude it cannot do the job, you may be abandoning a capable model prematurely. Teams that invest in elicitation routinely discover that models can handle tasks they initially dismissed, saving the cost of switching providers or building workarounds.

Common confusion: Elicitation uncovers capabilities that already exist in the model's weights. Fine-tuning can be part of the process, and the core goal is always surfacing what the model already knows how to do.