Documentation Hub Pattern

A documentation hub organizes knowledge into independent, domain-specific collections — each with its own navigation, content, and routing — while providing AI-consumable output via the llms.txt standard. The result is a single site that serves both human readers and LLM agents.

Field Notes itself is a working implementation of this pattern.

Multi-Collection Architecture

Each knowledge domain gets its own Fumadocs collection. Collections are defined in source.config.ts and loaded independently:

// source.config.ts
export const designSystem = defineDocs({ dir: 'content/design-system' });
export const principles = defineDocs({ dir: 'content/principles' });
export const claude = defineDocs({ dir: 'content/claude' });
export const platform = defineDocs({ dir: 'content/platform' });

Each collection maps to its own loader with an independent base URL:

// lib/source.ts
export const designSystemSource = loader({
  baseUrl: '/design-system',
  source: designSystem.toFumadocsSource(),
});

This produces independent sidebars, page trees, and search indexes per domain. Domains share a common navbar but are otherwise self-contained.

Navigation via meta.json

Each content directory uses a meta.json file to control sidebar ordering and grouping:

{
  "title": "Domain Name",
  "pages": [
    "index",
    "---Group Label---",
    "page-one",
    "page-two"
  ]
}

• Filenames are without extension ("my-page" not "my-page.mdx")
• Separator syntax: "---Label---" creates visual section dividers
• Pages appear in the order listed

Content Conventions

All content files are MDX with two required frontmatter fields:

---
title: Page Title
description: One-line summary that powers search and llms.txt.
---

• Lowercase kebab-case filenames
• Each domain has an index.mdx as its landing page
• Cross-reference within a domain using relative links
• Cross-reference across domains using absolute paths (e.g., /principles/answer-first)

llms.txt Generation

The llms.txt standard makes documentation sites AI-consumable. A build-time script reads all MDX files, strips JSX, and outputs two files:

The generation script runs automatically on npm run build. It parses frontmatter for titles and descriptions, strips import statements and JSX components from the body, and outputs clean markdown.

An LLM agent can fetch llms.txt to understand what's available, then fetch specific pages — or fetch llms-full.txt for one-shot ingest of all documented patterns.

Using as AI Context

To give another project's AI agent access to all documented patterns, add to that project's CLAUDE.md:

## Reference

See https://your-site.com/llms-full.txt for design engineering patterns
covering token systems, sprint frameworks, platform architecture, and AI workflows.

The agent reads the URL and has instant context on every pattern — no manual copy-pasting, no stale documentation.

CLAUDE.md as Self-Documentation

The hub's own CLAUDE.md follows the CLAUDE.md Playbook:

• Tier 1: What the project is, how to build it, key commands
• Tier 2: Compounding knowledge earned from real mistakes (frontmatter requirements, meta.json syntax, static export constraints)
• Git conventions: Branch prefixes, commit format
• Deployment: How and where the site deploys

This makes the CLAUDE.md both a practical reference for contributors and a demonstration of the playbook in action.

Why This Matters

The documentation hub pattern embodies several core principles:

• Distill, Don't Repeat (Synthesis & Validation) — Knowledge is synthesized from project repos into concise pages, not raw notes
• Answer First — Frontmatter descriptions ensure the key point is always visible before reading the full page
• Compounding Engineering — Content grows from real project needs, not speculative documentation
• Verification Over Instruction — The build process verifies all pages compile, all links resolve, and search indexes generate correctly