Comment by sungho_

How about adding these texts and reactions to LLM's context and iterating to improve performance? Keep doing it until a real person says, 'Yes, you're good enough now, please stop...' That should work.