The Intent tab picks one of 9 values. Under the hood every intent shares the exact same execution path — the only thing that changes is one sentence prepended to the system prompt. There's no other functional branching by intent, with one exception noted below. Pick the intent whose sentence matches what you're asking the model to do; don't expect deeper behavioral differences that aren't there.
The 9 intents
- classify — "Classify the provided data into one of the specified categories." Pair with
outputFormat: category. - extract — "Extract the requested information from the provided data." The one intent with special handling — see below.
- summarize — "Provide a clear, concise summary of the provided data."
- generate — "Generate content based on the provided data." The general-purpose choice — also what RAG-style grounded answers typically use.
- decide — "Analyze the provided data and make a decision." Pair with
outputFormat: boolean. - translate — "Translate the provided text accurately."
- analyze — "Identify patterns, trends, and anomalies in the provided data."
- validate — "Check the provided data against the specified rules and return specific error messages for each violation."
- moderate — "Review content for policy violations and flag inappropriate material."
extract is the one exception
When intent is extract AND outputFormat is json AND you've declared outputFields, the user message is restructured to lead with an explicit "Fields to extract:" list (each field's name and description) before the data. Every other intent just sends "Data:" followed by your context — the field list is extract-specific.
Your own Instructions field (the free-text box under Intent) is inserted between the intent sentence and the output-format instruction — this is where task-specific guidance belongs, since the intent itself only sets a general persona.
No dedicated "compare" intent
To compare options, feed the candidates into Context and use decide or classify to pick between them in one call — or run ai_call once per candidate and combine results in a downstream condition/formula step.