This is why I spent a lot of time building a design system first, that I can then re-use in my projects. My system is built out of properly CSS tokenized and organized Vue components, and the agent has instructions that it cannot create HTML that isn't made of components - and that if it needs a different UX (which I will have designed with it), then it needs to be built as a component.
What's nice is that you can then write linters for rando CSS and component violations most of the time.
Even then, you sometimes still have to beat it down when it goes rogue - but it does tend to bring my daily frustration down a bit. (Along with a process during design to preview these layouts before they get to a coding agent.)
And once you have a design system that suits your needs, its easy to add in as skills/scripts to any new project. I keep a github repo with all of these bits and bobs, with an agents.md that teaches the agent how to use the repo.
But in short: if the LLM keeps screwing it up, build tools and define it once - stop trying to prompt your way out of it every time.
itsdavesanders · · focus · HN ↗
What's nice is that you can then write linters for rando CSS and component violations most of the time.
Even then, you sometimes still have to beat it down when it goes rogue - but it does tend to bring my daily frustration down a bit. (Along with a process during design to preview these layouts before they get to a coding agent.)
And once you have a design system that suits your needs, its easy to add in as skills/scripts to any new project. I keep a github repo with all of these bits and bobs, with an agents.md that teaches the agent how to use the repo.
But in short: if the LLM keeps screwing it up, build tools and define it once - stop trying to prompt your way out of it every time.