# LLM resources

What to give an AI assistant so it writes correct Zazz markup.

Source: https://www.zazz.sh/docs/ai/llm-resources/

No model has current Zazz (0.5 and later) in its training data yet, and 0.4 looked different (classes, `data-variant`). An assistant needs the rules in its context. These are the files to hand it, from shortest to longest.

## Plain-text endpoints

- **[/llms.txt](/llms.txt)**: a one-screen orientation with links to every page as Markdown.
- **[/llms-full.txt](/llms-full.txt)**: every docs and API page in one file, tables included.
- **Any page as Markdown**: the `index.md` next to it: `/api/primitives/button/index.md`.

## The authoring rules

`AUTHORING.md` is the one-page source for the markup rules: identities, presets, slots, utilities, modifiers, hooks, layout, typography, theme. It ships inside `@zazz-ui/core` (`node_modules/@zazz-ui/core/AUTHORING.md`, or [on jsDelivr](https://cdn.jsdelivr.net/npm/@zazz-ui/core/AUTHORING.md)) and is what every other resource below points at. Paste it, or link it, before asking for any markup.

## Agent skills and rules

Two skills install into Claude Code, Cursor, Codex, and the other agents the [skills CLI](https://skills.sh) supports:

```bash
npx skills add dereknelsen/zazz-ui
```

- **`zazz`**: the design system: how to install and update it, the VS Code extension, its markup model and design guidance, references for tokens, components, and APIs, and where each answer lives on this site.
- **`install-zazz-in-legacy-app`**: for adding Zazz to an older app whose CSS must keep working. It interviews you one round of decisions at a time (delivery, rollout, layering, scripts, CSP, conflicts), each question with a recommended answer, then layers the old CSS and writes a scoped baseline that closes the gaps the Zazz reset opens.

Add one with `--skill zazz` or `--skill install-zazz-in-legacy-app`. The `zazz` skill says what the Cursor and Copilot rules files in the repo (`.cursor/rules/zazz-markup.mdc`, `.github/instructions/zazz-markup.instructions.md`) say: read `AUTHORING.md`, copy a primitive's anatomy from its page (`/api/primitives/<name>/index.md`) instead of reconstructing it, use tokens, and never use 0.4 syntax.

## The editor data

For an assistant that can read JSON, `editor/zazz.html-data.json` lists every `data-ui` token and every `data-<name>-*` attribute with its allowed values, and `editor/zazz.css-data.json` lists every utility, tier, and hook. They are generated from the kit, so they match the version they ship with: `node_modules/@zazz-ui/core/editor/`, or `https://cdn.jsdelivr.net/npm/@zazz-ui/core@<version>/editor/` (the CLI doesn't vendor them). `editor/zazz.language-data.json` adds each primitive's summary, hook defaults, and resolved tokens.

## Checking the output

Whatever the assistant writes, run it through the same audit a human gets: `<ui-debug>` in the page, or `eslint "**/*.html"` with the [plugin](/docs/tooling/editor-and-linting/). A prompt that ends with "then fix every `[ui-debug]` warning" closes the loop.
