Feature Request: Support prefill for Claude

[nissa-seru]

Prefill enables pre-population of the first part of the model's input, such that the model will continue responding as if it wrote the prefilled content. This enables chain-of-thought and other techniques to be used with very high compliance. For example, providing a psuedo-XML schema such as:

<schema name="response_process">
    <required_fields>
<field name="understand_request"></field>
 <field name="plan_response"></field>
<field name="respond_to_user"></field>
</required_fields>
</schema>

and prefilling with:

<response_process>
<understand_request>

works extremely well and is considerably more compact than most chain-of-thought prompting techniques.