Reducing AI-sounding output in a mostly stateless multimodal commentary bot

Prompting
1

I’m building a low-latency multimodal commentary bot with cloud API calls.

Per request, the model mainly gets:

  • the current screenshot
  • a short system prompt
  • a short user prompt

I’m deliberately avoiding long multi-turn history because latency and token cost matter.

The target output is short live-commentary style text, similar to viewer reactions:

  • casual
  • fragmented
  • frame-aware
  • varied in tone
  • not assistant-like

What I already tried:

  • persona/style rotation
  • explicit “no AI tone / no summary / no customer-support tone” prompt instructions
  • short JSON output contracts
  • exact/fuzzy dedup
  • dropping stale replies when the scene has already changed

The remaining problem is that the output can still feel too AI-written:

  • too clean
  • too balanced
  • too intentional
  • not enough short-term continuity across nearby moments

For people building with the API, what has worked best here?

Questions:

  1. How would you preserve short-term continuity without sending long history every turn?
  2. Is a small explicit state object usually better than summary text?
  3. What prompt patterns have actually helped reduce “AI tone” in short reactive outputs?
  4. At this point, is the bigger lever prompt design, decoding/sampling, or state architecture?
2

For me it was clear definition of what “human” should sound like.

Then taking human samples and converting then to “neutral” tone.

Then fine-tuning one model on reverse of the operation above (neutral-human, makes “human” skin).

Then fine-running another model on human-neutral (makes “neutral” skin).

Then running the operations like this:

  1. Generate “human” response (the definition of how goes in here)
  2. Use the neutralizer skin to get the base statement.
  3. User the “human” skin to get the final result from the base statement.