Prompt optimization and output formats: what the research shows.
Prompt engineering is writing a prompt by hand. Prompt optimization searches for a better one automatically. The shape of a reply can also change what a model gets right, and so can asking more than once or asking more than one model. We keep 31 published methods and studies on these, and test what applies on your questions.
31 published methods and studies, by what they change
Tools that search for a better prompt.
Each takes a task and a way to score it, tries prompts, and keeps what scores well.
The format of a reply can move the score, or not.
Some studies find a large effect and some find none. It depends on the model and the task.
Ways to get more right from the same model.
Each one spends more tokens, or answers fewer questions, to get more of its answers right.
Combining models, not choosing one.
Models get different questions wrong. These look for a gain from using more than one.
Reusing the start of a prompt.
Vendors charge less for input they have seen recently, and serve it faster. What counts is what stays the same at the start of each request.
So we test it on your questions.
Published results come from other models and other tasks, and they disagree. So prompts and formats are part of every run. We try plain and expert system prompts and prompts tuned with DSPy, GEPA and MIPROv2, against replies as JSON, as plain text and with the reasoning first. The report says which to use for which sub-task. See how it works for frontier models and open models.