How Vercel’s agents build on-brand pages with design.md →
The problem we found was that while the prompt described our visual language just fine, every model reading it interpreted that description differently, generating vastly different pages from the same guidance.
I like Vercel's approach of running the prompt twice, once with design.md and once without, to compare the impact design.md is actually having.
Right now most of us are still figuring out how to maintain brand consistency across multiple outputs using a single provider of choice (Anthropic/OpenAI). As it becomes more common for orgs to run multiple models—for different tasks and/or teams—we're going to need more mature ways to evaluate how effectively each model is able to adhere to guidelines.