Website Roaster.

End of run

A roast of the site that’s mean, accurate, and useful in that order. Critique that actually sticks because it’s funny, the kind a designer screenshots and pins above their monitor instead of skimming and ignoring like every other “audit report.”

74 seconds per roast, 8 cents a serve.

Challenges it solves

Generic audits get tuned out. The whole point of the persona is that the feedback lands. Wrapping real UX/copy/IA critique in a strong voice gets read where a polite bullet-list doesn’t.

One page isn’t the site. Most “AI website analyzers” judge a homepage in isolation. This one extracts internal links and scrapes the subpages too, so the roast can point at where the nav lies, where the pricing page contradicts the homepage promise, where the about page tells a different story than the product page.

Hold the voice. The hard part of any persona agent is keeping the voice tight without sacrificing accuracy. Claude is given the room to be brutal and theatrical, but every claim has to map back to something actually present on the scraped pages.

Under the hood

Start → Collect Website Details → Scrape Homepage → Extract Internal Links → Process Internal Links → Scrape Subpages → Analyze Website → Generate Roast → Display Roast → End. Single model (Claude 3.7 Sonnet) for the analysis and voice; the scraping and link-processing happen in deterministic nodes around it.