Prompt Conversion Criteria

Four conditions a prompt must satisfy at once to turn a user's problem into a completion that actually solves it.

Converting the user's problem into the model domain is the crux of prompt engineering: craft a prompt whose completion addresses the user's problem, satisfying four criteria simultaneously.

  1. Resemble training-set content — the Little Red Riding Hood principle: the prompt should closely resemble a document type the model has plausibly seen before.
  2. Include all relevant information — sometimes the user's raw input suffices; sometimes the application must pull in preferences, calendars, availability, news, or other external data. The two challenges are finding all possibly relevant content and finding the best of it — over- saturating the prompt with loosely relevant material distracts the model and degrades completions, the same problem the Chekhov's gun retrieval fallacy names for retrieval specifically. Content must also be arranged into a well-formatted, logical document, not just dumped in.
  3. Lead the model to a solution, not elaboration — the prompt must condition the model to actually address the problem rather than merely restate or elaborate on it. This is trickier for completion models, which need an explicit signal to switch from problem-poser to problem-solver; it is largely solved for chat models, which are fine-tuned to produce a helpful assistant reply automatically.
  4. Ensure the model reaches a stopping point — see completion stop-condition design for how this is engineered, particularly for completion models where nothing in the API tells the model a turn has ended.

Chat APIs largely satisfy criteria 1, 3, and 4 automatically, because the chat template resembles fine-tuning data, the model is conditioned to be helpful, and it stops at the end of its own message. The prompt engineer remains fully responsible for criterion 2 regardless of model type, and still needs to shape the system message, transcript, and tool definitions so the model can reach a genuinely successful stopping point under criteria 3 and 4.

Sources

  • Prompt Engineering for LLMs (John Berryman), ch. 4