Say it in your own words. Cannon scores it against every desk's vocabulary, picks a model for the shape of the request, and shows you both before anything runs.
Each desk has its own prompt, tools and retrieval scope. None can read another's data. If none of them fits, describe what you need above and Cannon will build one.
A pick moves that model to the front of the fallback chain. The rest stays behind it.