Series

Prompt Engineering Research

We have evaluated how different prompts really affect model performance across many domains and 5 different models.

AI Format Wars: Does the Shape of Your Prompt Matter? (1,080 Evals Later)
We tested 5 frontier models across 1,080 evaluations. JSON won. Short prompts won. Everything you thought you knew about prompt engineering is wrong.
Mar 22, 20265 min read22
What We Learned From Analyzing 28,000 Production AI System Prompts
80% start with "Act as an expert." 62% contain contradictory constraints. 14% emotionally threaten the AI. Here's the data.
Mar 24, 20262 min read8