Imported from mdocui/mdocui (SKILL.md). Install upstream with npx skills add mdocui/mdocui. Copyright stays with the author.

mdocUI — Implementation Skill for AI Agents

This file teaches AI coding agents (Claude Code, Cursor, Copilot, etc.) how to implement with mdocUI. Install in Claude Code: copy this file to your project's .claude/ directory or reference it in CLAUDE.md.

What is mdocUI

mdocUI is a generative UI library for LLMs. It uses Markdoc {% %} tag syntax inline with markdown prose. The LLM writes natural markdown AND drops interactive UI components in the same stream.

npm packages:

• @mdocui/core — streaming parser, component registry, prompt generator
• @mdocui/react — React renderer, 24 theme-neutral components
• @mdocui/cli — scaffold, generate system prompts, preview

Installation

pnpm add @mdocui/core @mdocui/react

Core Concepts

1. The LLM generates this syntax

Here are the results.

{% chart type="bar" labels=["Q1","Q2","Q3"] values=[100,150,200] title="Revenue" /%}

{% callout type="info" title="Note" %}
Revenue is trending up.
{% /callout %}

{% button action="continue" label="Show details" /%}

Self-closing tags: {% tagname attr="value" /%} Body tags: {% tagname %} content {% /tagname %}

2. System prompt is auto-generated

import { ComponentRegistry, generatePrompt, allDefinitions, defaultGroups } from '@mdocui/core'

const registry = new ComponentRegistry()
registry.registerAll(allDefinitions)

const systemPrompt = generatePrompt(registry, {
  preamble: 'You are a helpful assistant.',      // your domain context
  groups: defaultGroups,                          // component organization
  additionalRules: ['Use charts for data'],       // domain-specific guidance
  examples: ['{% chart type="bar" ... /%}'],      // expected output format
  verbosity: 'default',                           // 'minimal' | 'default' | 'detailed'
})
// Pass systemPrompt as the system message to your LLM

allDefinitions and defaultGroups are exported from both @mdocui/core and @mdocui/react.

The prompt merges two layers:

• Library layer (auto): tag syntax, component signatures, composition rules
• App layer (yours): preamble, rules, examples, groups

The verbosity option controls prompt size: minimal (compact, tag names only), default (signatures + basic rules), detailed (full descriptions + examples).

3. Render with React

import { Renderer, defaultComponents, useRenderer, createDefaultRegistry } from '@mdocui/react'

const registry = createDefaultRegistry()

function Chat() {
  const { nodes, isStreaming, push, done, meta } = useRenderer({ registry })

  // Feed LLM chunks: push(chunk)
  // When stream ends: done()
  // useRenderer automatically batches renders to at most one per frame (~60fps)

  return (
    <Renderer
      nodes={nodes}
      components={defaultComponents}
      registry={registry}                // enables Zod prop validation after stream ends
      isStreaming={isStreaming}
      meta={meta}                        // enables shimmer placeholders for pending tags
      contextData={{ userId: '123' }}    // arbitrary data passed to all components
      classNames={{ button: 'bg-blue-600 text-white', card: 'border-zinc-700' }}
      onAction={(event) => {
        if (event.action === 'continue') sendMessage(event.label)
      }}
      onError={(event) => {
        console.error(`Component <${event.componentName}> failed:`, event.error)
      }}
      renderProse={(text, key) => <ReactMarkdown key={key}>{text}</ReactMarkdown>}
      renderPendingComponent={(pendingTag) => <MyShimmer tag={pendingTag} />}  // or null to disable
    />
  )
}

4. Style with classNames (not in the library)

The library renders theme-neutral HTML. The app controls all colors:

<Renderer
  classNames={{
    button: 'bg-blue-600 text-white rounded-lg',
    card: 'border border-zinc-200 rounded-xl shadow-sm',
    callout: 'border-l-4 border-blue-500 bg-blue-50',
    chart: 'my-chart-class',
  }}
/>

Or override via CSS targeting data-mdocui-* attributes:

[data-mdocui-chart-bar] { background: #3b82f6 !important; }
[data-mdocui-button] { background: #0f172a; color: white; }

5. Custom components

Override any component or bring your own:

const myComponents = {
  ...defaultComponents,
  button: MyButton,      // swap one
  card: MyShadcnCard,    // use shadcn
}

<Renderer components={myComponents} />

Every component receives ComponentProps:

interface ComponentProps {
  name: string
  props: Record<string, unknown>
  children?: React.ReactNode
  className?: string
  onAction: ActionHandler
  isStreaming: boolean
}

Renderer Props

interface RendererProps {
  nodes: ASTNode[]                    // parsed AST from useRenderer
  components?: ComponentMap           // component map (default: defaultComponents)
  registry?: ComponentRegistry        // enables Zod prop validation after streaming
  onAction?: ActionHandler            // single callback for all interactions
  onError?: (event: ComponentErrorEvent) => void  // catch component render errors
  isStreaming?: boolean               // true while LLM is still streaming
  meta?: ParseMeta                    // parser metadata (enables shimmer for pending tags)
  contextData?: Record<string, unknown>  // arbitrary data passed to all components
  renderProse?: (content: string, key: string) => React.ReactNode
  renderPendingComponent?: ((pendingTag?: string) => React.ReactNode) | null  // custom shimmer, or null to disable
  classNames?: Record<string, string> // per-component Tailwind/CSS classes
}

interface ComponentErrorEvent {
  componentName: string               // name of the component that failed
  error: Error                        // the thrown error
  props: Record<string, unknown>      // props that were passed to the component
}

Available Components (24)

Layout

• stack — flex container (direction, gap, align)
• grid — CSS grid (cols, gap)
• card — bordered container (title, variant)
• divider — horizontal line
• accordion — collapsible section (title, open)
• tabs — tabbed container (labels, active) — children: tab only
• tab — tab panel (label)

Interactive

• button — action button (action, label, variant) — disables after click
• button-group — button row (direction) — children: button only
• input — text field (name, label, placeholder, type)
• textarea — multi-line input (name, label, placeholder, rows)
• select — dropdown (name, label, options, placeholder)
• checkbox — toggle (name, label, checked)
• toggle — switch (name, label, checked)
• form — form container (name) — children: input, textarea, select, checkbox, toggle, button

Data

• chart — visualization (type: bar|line|pie|donut, labels, values, title)
• table — data table (headers, rows, caption)
• stat — metric display (label, value, change, trend: up|down|neutral)
• progress — progress bar (value, label, max)

Content

• callout — alert block (type: info|warning|error|success, title)
• badge — inline tag (label, variant)
• image — image (src, alt, width, height)
• code-block — code block (code, language, title)
• link — clickable link (action, label, url)

Composition Patterns

Nest layout components to build rich dashboards:

{% card title="Dashboard" %}
  {% grid cols=3 %}
    {% stat label="Revenue" value="$12K" change="+8%" trend="up" /%}
    {% stat label="Orders" value="284" change="+12%" trend="up" /%}
    {% stat label="AOV" value="$43" change="-2%" trend="down" /%}
  {% /grid %}
{% /card %}

{% card title="Trends" %}
  {% chart type="bar" labels=["Mon","Tue","Wed"] values=[100,150,200] /%}
{% /card %}

{% tabs labels=["By Region","By Channel"] %}
  {% tab label="By Region" %}
    {% table headers=["Region","Revenue"] rows=[["US","$8K"],["EU","$4K"]] /%}
  {% /tab %}
  {% tab label="By Channel" %}
    {% chart type="pie" labels=["Direct","Social","Email"] values=[45,30,25] /%}
  {% /tab %}
{% /tabs %}

Action Handling

onAction={(event) => {
  // event.type: 'button_click' | 'form_submit' | 'select_change' | 'link_click'
  // event.action: string identifier
  // event.label: button/link text
  // event.formState: form field values (on submit)
  // event.params: extra data (e.g. { url } for links)

  if (event.action === 'continue') sendMessage(event.label)
  if (event.action.startsWith('submit:')) sendMessage(JSON.stringify(event.formState))
  if (event.action === 'open_url') window.open(event.params.url, '_blank')
}}

Error Handling

onError={(event) => {
  // Fires when a component throws during render
  console.error(`Component <${event.componentName}> failed to render`)
  console.error('Error:', event.error)
  console.error('Props:', event.props)
  
  // After a successful render, the error boundary shows the last valid output.
  // On first-render failures, a simple error message is shown instead.
}}

Streaming API Route (Next.js)

import { ComponentRegistry, generatePrompt, allDefinitions, defaultGroups } from '@mdocui/core'

const registry = new ComponentRegistry()
registry.registerAll(allDefinitions)
const systemPrompt = generatePrompt(registry, { preamble: '...', groups: defaultGroups })

export async function POST(req: Request) {
  const { messages } = await req.json()
  // Pass systemPrompt as system message to any LLM
  // Stream response as text/plain — the client parser handles the rest
}

Common Patterns

Grid whitespace fix

The streaming parser emits whitespace prose nodes between component tags. Inside a grid, these take up grid cells and break layout. Fix by skipping whitespace-only prose:

renderProse={(text, key) => {
  if (!text.trim()) return null
  return (
    <div key={key}>
      <SimpleMarkdown content={text} dataKey={key} />
    </div>
  )
}}

Or via CSS as a fallback:

[data-mdocui-grid] > div:has(> [data-mdocui-prose]:only-child) {
  display: none !important;
}

Product card layout (PLP/PDP)

Compose existing components for shopping-style product cards:

{% grid cols=3 %}
{% card %}
{% image src="https://example.com/shoe.jpg" alt="Running Shoe" /%}

**Running Shoe Pro**

**$149.99** · 4.8★ · {% badge label="In Stock" variant="success" /%}

Lightweight mesh with responsive cushioning.

{% button action="continue" label="View Running Shoe Pro" variant="primary" /%}
{% /card %}
{% /grid %}

Use prose for price/rating (not stat — it renders too large for product cards). Include the product name in button labels so follow-up messages have context.

Styling via CSS data attributes

All components render data-mdocui-* attributes. Style them globally without classNames:

[data-mdocui-card] { border-radius: 12px; box-shadow: 0 1px 3px rgba(0,0,0,0.04); }
[data-mdocui-stat][data-trend="up"] > div:last-child { color: #16a34a; }
[data-mdocui-stat][data-trend="down"] > div:last-child { color: #dc2626; }
[data-mdocui-callout][data-type="warning"] { background: #fffbeb; border-left-color: #d97706; }
[data-mdocui-badge][data-variant="success"] { background: #dcfce7; color: #15803d; }

Disable shimmer flicker

The built-in shimmer flickers during streaming because pendingTag toggles rapidly as each tag opens/closes. Disable it and use a simple typing indicator instead:

<Renderer renderPendingComponent={null} />

Built-in Prose Rendering (SimpleMarkdown)

SimpleMarkdown is a zero-dependency markdown renderer for prose nodes. Supports:

• Inline: bold (**), italic (*), bold+italic (***), strikethrough (~~), inline code, links
• Block: headings (h1-h3), unordered lists (-/*), ordered lists (1./1)), paragraphs

Headings and lists are detected line-by-line — no blank line separation required.

Override with a full markdown renderer if you need GFM tables, images, or nested lists:

<Renderer
  renderProse={(text, key) => <ReactMarkdown key={key}>{text}</ReactMarkdown>}
/>

To skip rendering whitespace-only prose (fixes grid layout issues with whitespace nodes):

renderProse={(text, key) => {
  if (!text.trim()) return null
  return <SimpleMarkdown content={text} dataKey={key} />
}}

Key Principles

1. LLM generates structure, client handles styling — never put colors in the system prompt
2. Prose + components in one stream — markdown flows naturally, {% %} tags add interactivity
3. Single source of truth — Zod schemas drive both validation AND prompt generation
4. Theme-neutral defaults — components use currentColor/inherit, app decides colors via classNames
5. One-shot interactions — buttons disable after click, forms lock after submit
6. Consecutive buttons auto-group inline — no layout work needed

Links

• GitHub: https://github.com/mdocui/mdocui
• npm: https://www.npmjs.com/org/mdocui
• Docs: https://mdocui.github.io
• Examples: https://github.com/mdocui/examples
• Live Demo: https://mdocui.vercel.app