Prefilling

Prefilling is an API-level technique that lets you supply the opening text of Claude's response by placing a final message in the conversation array with the role set to 'assistant'. Instead of waiting for Claude to choose how to begin its reply, you inject the first characters yourself—for example, an opening curly brace to force JSON output, or a persona marker to lock a role-play character. Claude then continues generating from exactly where your prefill ends, making it a precise way to control output format, skip conversational filler, and maintain character consistency. The feature works by appending a dictionary with 'role': 'assistant' and your desired opening text as 'content' to the messages list in an API call. Claude treats that text as if it had already written it and completes the response from that point forward. This bypasses any preamble Claude might otherwise add (such as 'Certainly! Here is…') and steers the model directly into a specific output trajectory before a single new token is generated. Prefilling is strictly an API feature—it is not available on Claude.ai. It is also subject to important model-version restrictions: starting with the Claude 4.6 family and Claude Mythos Preview, prefilling is explicitly disabled at the API level and returns a 400 error. On supported models it remains a useful low-overhead technique, but Anthropic now recommends structured outputs and system-prompt instructions as the primary approach for newer models.