One URL for any AI

When a developer asks an AI coding assistant (Claude Code, Cursor, Copilot, Gemini, Windsurf) to build a website with @webhouse/cms, they can now just say:

"Use https://ai.webhouse.app to build it."

The AI fetches that URL and gets a self-contained Step 0–9 walkthrough: verify environment → scaffold → plan with the user → edit cms.config.ts → create content → wire rendering → SEO → deploy → hand back. The page is served as text/markdown, with zero visual chrome — purpose-built for LLM consumption.

All 21 deep-dive modules live at ai.webhouse.app/ai/{slug} if the AI needs depth on a specific topic.

What ships at ai.webhouse.app

All responses set X-Robots-Tag: noindex — the AI site is not indexed by traditional search engines. It's there for machines.

Why a dedicated AI site?

The 21 modules already exist on GitHub raw, but that's not enough:

1. Discoverability — one pretty URL is easier to remember and type than a deep GitHub path.
2. Rate limits — GitHub raw throttles at 60 requests/hour unauthenticated. A fresh AI session that fetches several modules hits that quickly.
3. Platform variance — not every AI coding tool has ergonomic URL-fetch tooling for GitHub paths, but essentially all of them can fetch a plain URL like ai.webhouse.app.
4. Self-guided entry point — the existing index is a module list. It assumes the AI already knows what @webhouse/cms is. The new /ai walkthrough is written for a blank AI session — "you are an AI, follow these steps, if you need X fetch Y".
5. Machine-readable endpoints — AI platforms increasingly expect llms.txt, structured manifest JSON, and versioning. GitHub raw markdown doesn't provide any of that.

The walkthrough (/ai)

The walkthrough is structured as a procedure, not a reference:

Step 0 — Verify environment
Step 1 — Scaffold the project (npm create @webhouse/cms@latest)
Step 2 — Understand the model (document JSON, cms.config.ts, field types)
Step 3 — Plan the site with the user (5 questions)
Step 4 — Edit cms.config.ts (kind + description required per collection)
Step 5 — Create starter content
Step 6 — Wire up rendering (Next.js / static / non-TS consumer)
Step 7 — SEO (if relevant)
Step 8 — Deploy
Step 9 — Hand back to the user

Troubleshooting — first-pass fixes
Deep-dive module index
Non-negotiable rules (8 rules from CLAUDE.md critical rules)

The AI can complete a basic build end-to-end without fetching any other module. Deeper modules are only for specific topics (i18n, complex SEO, non-TS backend).

How to use it

In any AI coding tool, paste:

Build me a website with @webhouse/cms. Start by fetching https://ai.webhouse.app

The AI will:

1. Fetch /ai and follow Step 0 (check node --version, ask where to scaffold)
2. Run npm create @webhouse/cms@latest (Step 1)
3. Explain the model and plan with you (Steps 2–3)
4. Edit cms.config.ts, create content, wire rendering (Steps 4–6)
5. Fetch specific deep-dive modules as needed (SEO, i18n, deployment)
6. Hand back a running local site with deploy hooks configured

No manual CLAUDE.md setup, no copy-pasting config from Stack Overflow, no "where should I start?" confusion.

Works with any AI platform

The entire site is plain text/markdown and application/json — every AI coding assistant understands it:

• Claude Code — WebFetch tool pulls the URL; the long context window handles the full walkthrough
• Cursor / Windsurf — paste the URL; the inline chat fetches it
• GitHub Copilot — chat supports URL context; fetch /ai into the prompt
• Gemini (IDE extensions) — URL-aware chat fetches the markdown
• Any future platform — as long as it supports URL fetching, it works

Machine-readable endpoints

For programmatic use (MCP servers, agent frameworks, CI pipelines):

# Full JSON manifest with all modules + descriptions + endpoints
curl https://ai.webhouse.app/ai/manifest.json

# Plain ordered module list
curl https://ai.webhouse.app/ai/index.json

# llms.txt standard (for LLM crawlers)
curl https://ai.webhouse.app/ai/llms.txt

Example manifest.json:

{
  "name": "@webhouse/cms AI Builder Site",
  "version": "0.1.0",
  "entry": "https://ai.webhouse.app/ai",
  "modules": [
    { "slug": "01-getting-started", "url": "...", "description": "New project, first setup" },
    ...
  ],
  "endpoints": {
    "walkthrough": "https://ai.webhouse.app/ai",
    "llms_txt": "https://ai.webhouse.app/ai/llms.txt",
    "manifest": "https://ai.webhouse.app/ai/manifest.json"
  }
}

Canonical source + fallbacks

If ai.webhouse.app is unreachable, the canonical markdown lives in the CMS monorepo and is served from GitHub raw:

https://raw.githubusercontent.com/webhousecode/cms/main/docs/ai-guide/index.md
https://raw.githubusercontent.com/webhousecode/cms/main/docs/ai-guide/01-getting-started.md
...

Every scaffolded project also ships with packages/cms/CLAUDE.md (slim index, ~180 lines) that references the same modules. So the AI has three redundant paths to the same content:

1. https://ai.webhouse.app (this site — primary)
2. GitHub raw (fallback, rate-limited)
3. Local packages/cms/CLAUDE.md in node_modules/@webhouse/cms (offline)

Dogfooding

The AI Builder Site itself is built with @webhouse/cms — it lives as a route group (src/app/ai/*) on docs.webhouse.app, which is entirely CMS-managed. One site, two audiences: /docs for humans, /ai for AI agents. Same content model, same deploy pipeline.

This page (the one you're reading) is a document in the docs collection, stored as JSON at content/docs/ai-builder-guide.json. It explains the companion AI site that sits next to it. Self-referential, fully editable via the admin UI, and available in both English and Danish.

Non-negotiable rules

The walkthrough reinforces eight rules that AI sessions must follow when building with @webhouse/cms:

1. Every collection MUST have kind and description
2. cms.config.ts MUST declare storage explicitly (default is SQLite)
3. NEVER name a collection site-settings, settings, config, admin, media, interactives
4. image-gallery values MUST be { url, alt }[] — never plain strings
5. Documents MUST have _fieldMeta: {}
6. Slug MUST match filename
7. Always filter by status === "published" when rendering
8. NEVER use CDN scripts (Tailwind, Bootstrap) in static builds — inline CSS only

These rules prevent the most common AI-generated mistakes we've seen across early builds.