Chat role markers

The prompt

<|system|> ... <|user|> ... <|assistant|> — most chat-tuned models are trained on these role boundaries. Wrapping instructions in the correct role tag gets far more reliable adherence than plain prose alone.

Expected result

Higher adherence to the instruction than the same text without role markers, since the model was fine-tuned to weight system and user turns differently.

Why it works

Most hosted APIs abstract this away with a messages array — these raw markers matter most when prompting an open-weight model directly.