How to keep an AI agent in character

The hard part of building character-based is often not the model itself, but stopping the agent from slipping back into a generic helpful assistant style. Soft instructions like “be warm but concise” are weaker than the model’s built-in habits. Clear rules and control mechanisms work better than adjectives.

A strict response length limit is the strongest tone control. Long replies let the default assistant style leak back in, while short replies make the character come through more clearly. A special mode should be the main , not an extra section added onto it.

If the agent must use a tool, tool_choice should be set as required instead of leaving optional.

Key points

  • can ignore gentle style instructions and return to a default assistant voice.
  • Hard rules work better than vague words.
  • Strict length limits help preserve character and reduce tokens.
  • A mode should be the main , not a side note.
  • Use required tool_choice when a tool action must happen.
Read original