Every new site starts the same way. Paste a long build prompt into Claude, read the output, then spend three hours patching what it forgot. The max-image-preview:large meta. The focus-visible styles. The <picture> fallbacks. The llms.txt. The Sitemap: line in robots.txt. Always a few of them. Often all of them.
I got tired of it. Single Site Gen is the fix.
It generates the prompt you'd write if you had every rule the Mega Analyzer, Site Analyzer, and E-E-A-T Audit already check. In your head, all at once. Paste the output, get a site back that scores 90+ on the first deploy.
What's in the prompt
Sixteen modules, all on by default, all toggleable. The interesting ones:
Crawl validation. The prompt tells the model to ship a robots.txt with no Disallow conflict against any indexable path, a Sitemap: line, and a sitemap.xml with a valid <urlset> root containing every page. Every page gets two or more internal outbound links so nothing becomes a dead end. These are the same checks the analyzer will run post-deploy.
Schema. Not just "add JSON-LD". The prompt names every type. Organization. Person. BreadcrumbList. FAQPage. Speakable. Article with the text|image|video requirement and the datePublished timezone rule. WebSite with SearchAction.
E-E-A-T. Three or more sameAs profiles including Wikidata and ORCID where they apply. rel="me" identity links. A Person schema with jobTitle, knowsAbout, alumniOf, worksFor.
Core Web Vitals. Width and height on every image. WebP/AVIF via <picture>. Critical CSS inline. LCP under 2.5s, CLS under 0.1. Stated as targets, not as hopes.
The other modules. Accessibility, security headers, voice, answer-engine optimization, syndication feeds, IndexNow, analytics, the .well-known suite. Follow the same pattern. Specific rules, not vibes.
Three sizes
Pick the one that fits the conversation. Mega gives you about 4,000 words and every rule spelled out. Quick is closer to 900 words, the essentials. Mini is a single paragraph. Useful when you're embedding the spec inside a larger chat.
The loop that matters
- Fill the form. Name, domain, stack, host, pages, colors, keywords.
- Uncheck any module you don't need.
- Copy the prompt. Paste into Claude.
- Deploy the result to Netlify or Vercel.
- Run the URL through the Mega Analyzer.
- Whatever the analyzer flags, fix. Redeploy.
Step 6 is the whole point. The model will miss one or two rules on any given run. The analyzer will name them. The gap closes in one iteration instead of three hours.
Why this exists
Every LLM-content tool on the market either writes vague prompts full of "best practices" with no verification, or it enforces its own opinionated boilerplate you can't edit. This tool encodes the exact rules our audit suite checks. So there's a round-trip. Prompt builds the site. Audit grades it. Gap feeds back.
That's the methodology Chapter 41 of The $97 Launch calls The Master AI Prompt. The book ships a static artifact. This version is editable, modular, and self-auditing. And it knows about every rule we've added since the book went to print.