Structurer les messages et les préférences

Beyond the prompt, you can shape the call: a system prompt, model hints, length, and how much past context to include.

Use systemPrompt to set the model’s role or tone for this one call, like telling it to be a terse code reviewer.

The messages array can hold several turns, letting you give the model a short back-and-forth as context, not just one line.

modelPreferences takes hints — substrings like claude-3-sonnet — that the client may match to an available model.

The client makes the final pick. A hint is a suggestion; the host may map it to an equivalent model from another provider.

List hints in order of preference. A broad fallback like a bare claude hint catches any Claude model if your first pick is missing.

Set intelligencePriority from 0 to 1. Higher values nudge the client toward more capable, smarter models for hard work.

Need low latency? Raise speedPriority toward 1 and the client leans on faster models, trading some depth for quickness.

Raise costPriority to prefer cheaper models. These three priorities are normalized hints, not hard rules.

An includeContext setting lets the request pull in context from MCP servers, so the model sees more than just your text.

Where supported, a temperature value shifts replies from focused and deterministic to more varied and creative.

One question on guiding model choice.

Shape calls with a systemPrompt, layered messages, and modelPreferences — hints plus cost, speed, and intelligence priorities. 🎛️