You write a system prompt that says, clearly, in the first line: plain text only, no markdown, no asterisks. The agent responds with bullet points and bold headers.
You try a different model. Same result. You move the instruction to the bottom of the prompt. Same result.
This is not a model capability problem. It is a prompt positioning problem.
Models weight recent context heavily. By the time the model generates its response, the system prompt is ancient history. Your formatting directive is a note left on a desk three days ago.
Inject the formatting constraint immediately before every response -- appended to the user message, not buried in the system prompt.
In your message preparation code, before building the messages payload:
if history and history[-1]["role"] == "user":
history[-1]["content"] = history[-1]["content"] + "\n\n[FORMATTING RULE: Plain text only. No markdown. No asterisks, no bold, no headers, no bullet points.]"
When the formatting rule is the last thing the model sees before it generates, it sticks. When it is in the system prompt and nowhere else, it is advisory.
Leave the formatting directive in your system prompt as well. The system prompt version sets the expectation. The per-message injection enforces it. Both together are more reliable than either alone.
Some models are more resistant than others. Reasoning-class models in particular tend to override formatting instructions because they are optimised to produce structured output. If you are running a reasoning model on a tier where plain text matters, the per-message injection is not optional -- it is the only thing that works.