Prompts
Installing the code gives you this implementation. The prompt gives you the technique — which is what you want when the component needs to be built against your own markup, your own state, or a framework that is not React.
Every component page carries one, ready to copy. Paste it into Claude Code, Codex, Gemini or any agent with write access to your repository.
What is in it
A prompt is not a description of the component. It is a specification, generated from the same metadata that drives this site, so it can never drift from the code shown next to it. Nine sections:
What it does
The effect in plain language, so the model builds the right thing before it writes a line.
How it works
The mechanism as ordered steps. This is the part that stops a model inventing a plausible but wrong approach.
Constants — use these values
Every tuned number with the reason it holds. Without this a model picks round numbers and the motion feels wrong in a way nobody can name.
API
The exact signature — prop names, types, defaults.
Reference implementation of the core
The load-bearing region of the real source, excerpted. Not the whole file: enough that the model matches the technique, little enough that it still writes code in your conventions.
Requirements
Dependencies, cleanup obligations, and reduced-motion behaviour stated as requirements rather than suggestions.
Known pitfalls — avoid these specifically
The mistakes this technique actually invites. Naming them up front is cheaper than reviewing them out later.
Notes
Anything else that changes the implementation — browser limits, ordering constraints, licence boundaries.
Definition of done
The observable test. The model knows when to stop, and so do you.
Using it
Copy from the component page, or download the same text as a markdown file and point the agent at the path — useful when the prompt is long enough that pasting it into a terminal is awkward.
Say what is different about your project
The prompt describes the technique completely and says nothing about your codebase. Add that yourself in a sentence or two — it is the difference between a component that works and one that fits.
Letting the agent choose the component
If you do not know which component you want, give the agent the index instead and let it pick. /llms.txt is one line per component; /llms-full.txt carries every brief inline, at the cost of being large.
Both are plain text and need no key. You can also read the registry itself at /r/registry.json, or browse the same catalogue by hand on the components page.