WiderGood

Stop writing prompts. Make the model write them.

You know your problem. You don't need to know how to phrase it for a language model — that part is the model's job, and it is measurably better at it than you are. Describe the job, make it write the prompt, then run it.

Not folklore: model-written instructions matched or beat human annotators on 19 of 24 tasks (Zhou et al., 2022), and self-optimized prompts beat human-designed ones by up to 50% on harder reasoning benchmarks (Google DeepMind, 2023).

Copy this prompt

Write me a comprehensive prompt to put into an LLM that will address this need: [describe the need in plain language, including who the output is for, what format you want, and anything that would make the result wrong]. Before you write it, ask me any questions you need answered to write a better one.

Bring a few areas that eat your team's time. Thirty minutes. If AI can't fix them, I'll say so.