Advice Conversation Document
Frame a prompt as a conversation where one party asks for help and the other provides it, the document type ChatML itself is built around.
A prompt plus its completion together form a document, and per the Little Red Riding Hood principle the best document type to write is one the model has plausibly seen many of in training. The advice conversation — one person asking for help, another providing it — is one such archetype: the model plays the advisor, and the application or user plays the asker.
It is the natural fit for chat models (OpenAI built ChatML around advice conversations specifically because they judged them the most universally useful and easiest to implement), but it's useful for completion models too. Advantages:
- Natural interaction — people already think in conversational terms; you can pose a question and take the model's continuation as the answer.
- Multiround interactions — complex interactions continue the same prompt with new question/answer turns, with your own application logic interleaved between queries.
- Real-world integration — the conversational shape suits multiround processes and integrating real-world tools, with either chat or completion models, via function calling.
Using this structure with a chat model gets you RLHF-driven instruction compliance for free; using it with a completion model lets you sidestep unhelpful RLHF traits (particular stylistic habits, content policing) while still getting the conversational shape. Completion models writing this archetype need an explicit transcript format chosen, since nothing enforces one automatically the way a chat API's template does.
Sources
- Prompt Engineering for LLMs (John Berryman), ch. 6