“I built a prompt that can make any prompt 10x better”
funbike · 5d ago · Edited 5d ago Instead, I have the LLM generate and evaluate the prompt for me: 1. I tell the LLM to "Generate a detailed prompt engineering guide. The audience is <role>". (Example role might be "book authors" or "software developers" or "customer support reps"). 2. I paste in 5 examples of how I want my prompt to work (few-shot input+output). 3. I instruct it to "Generate a prompt that could have generated the examples' outputs, and include a better set of examples." Submit. 4. In a new chat, I instruct it to "Generate a detailed prompt evaluation guide. The audience is <role>". 5. I paste the new prompt and tell it to "Evaluate the prompt". 6. I tell it to "Generate 3 improved alternative prompts". 7. I pick the best one, and edit it as necessary. A benefit of this method is the LLM's own weights influence how the prompt is generated and evaluated. The prompt will be better than anything you could write. It's best to use a model in the same family as you'll use the prompt for. This works incredibly well and is very easy to do.
— GREG ISENBERG · original post ↗