Site Slap Devlog #1 V0 (the ai gave arngren.net a 42)

TL;DR

• built the full backend: screenshot, vision model, score + roast + fixes
• wired it to a live ui and deployed on Vercel
• tuned the prompt

---

the idea

paste a url. get a score. get roasted.

that’s the whole thing. no seo metrics, no performance audit. just a vision model looking at your site the way a person would, and being honest about it.

---

how the backend works

three steps, one route.

first, microlink takes a screenshot of the live site. it’s a simple GET request, no chromium to manage, no headless browser on the server. just a url in, a screenshot url back.

then that screenshot url goes straight to Groq’s vision model, llama 4 scout. it sees what the site actually looks like, not the html. the prompt tells it to score on visual vibe only, and return raw JSON with a score, a verdict, a roast, and three fixes.

then /api/slap hands that JSON back to the client.

it works.

---

the markdown fence bug

linear.app broke it on the first test. Groq decided to wrap the JSON in a code block even though the prompt said not to. one line fix: strip the fences before parsing. linear scored fine after that.

---

what i learned

the model is too generous by default.

arngren.net is one of the most visually chaotic sites on the internet. it got a 42. that’s “meh” territory. the prompt said “0-30 is painful” but the model hedged toward the middle anyway.

turns out you have to anchor the scale with specific examples and tell it to actually use the low end. added “arngren.net level” next to the 0-15 range and “be harsh when it’s bad” as a direct instruction. scores tightened up.

the prompt is the model’s personality. treat it like code.

---

Tip of the Day

go slap your own site before someone else does. you already know what’s wrong with it.