Controlling model output

Influencing LLM Output: Advanced Techniques

**I. Pre-filling Assistant Messages**

• Purpose: To strongly steer or direct Claude's response toward a specific stance, conclusion, or direction.
• Mechanism:
• Assemble a message list containing the user prompt.
• Add a final message designated as an "Assistant Message" (a Q&A format).
• This Assistant Message contains a partial, directional statement (e. g. , "Coffee is better because... ").
• How it Works: Claude assumes that the content in the Assistant Message was authored by it. It uses this partial text as a starting point and continues generating the rest of the response, effectively justifying the pre-filled text.
• Key Consideration: The pre-filled text is not guaranteed to be a complete sentence. The model continues generating text from the very end of the provided string, which may require the user to stitch the final output together.
• Application: Can be used to force a specific conclusion (e. g. , favoring tea) or a neutral/negative conclusion (e. g. , "neither is very good because... ").

**II. Stop Sequences**

• Purpose: To force Claude to immediately halt response generation as soon as it produces a specific string.
• Mechanism:
• Provide a list of stop sequences as an argument to the chat function.
• How it Works: When the model generates the exact sequence provided in the stop list, the entire generation process stops immediately.
• Application: Useful for controlling the length or format of the output (e. g. , stopping a counting sequence at a specific number).
• Note: If the goal is to stop before a sequence is fully generated, the stop sequence must include surrounding characters (e. g. , using "comma space five" to stop before the number five).

Takeaways

• Pre-filling Assistant Messages steers LLM output by providing a partial, directional statement that the model assumes it authored and continues to justify.
• Stop Sequences force the LLM to immediately halt response generation as soon as it produces a specified string.
• The pre-filled text may not be a complete sentence, requiring the user to stitch the final output.
• To stop generation before a sequence is fully produced, the stop sequence must include surrounding characters.