How can I monitor hundreds of custom prompts across several LLMs?
Module
Bikma's Question Analysis module builds the prompt set from the way people actually ask and weights it with the Prompt Probability Index, instead of rewriting keywords as questions.
How the prompt set is built
Bikma starts from the market's explicit and latent demand, segmented into brand and non-brand clusters and organised by topic. Each intent is sampled with several wordings, because a model's answer changes more when the question is rephrased than when it is repeated. A typical set for a B2B vertical is 60–120 well-chosen questions, and it is editable: manually added questions stay flagged so they do not contaminate the historical series.
The Prompt Probability Index
Bikma assigns each question a stated probability of being asked, not an observed volume, using an inference model validated with the University of Pisa and the University of Insubria. The metric is aggregated by intent rather than by single phrase, so the weight reflects how much a question really counts in the market.
Why it matters
The quality of the question set decides the quality of everything else: measuring the wrong questions, or weighting them badly, produces precise but useless numbers.