Question

You are building a product-review analysis feature that must return each result as JSON with fields 'sentiment', 'topics', and 'summary' so a downstream service can parse it reliably. In your testing, Claude sometimes wraps the JSON in conversational text like 'Sure! Here is the analysis:' which breaks the parser. Using the Messages API, what is the most effective way to get consistently parseable structured output?

Accepted Answer

In the system prompt, specify the exact JSON schema required and prefill the assistant turn with an opening '{' to force the response to begin as JSON

Answer

Set temperature to 0 and hope the model stops adding conversational text

Answer

Increase max_tokens so the model has enough room to include both the explanation and the JSON

Answer

Send the request without a system prompt so the model has no instructions to add commentary