I spent a day trying to break Readdy AI @ReaddyAI. Here's what actually broke. Started clean — a portfolio site for a graphic designer, light background, standard layout. 85% match to the brief in one prompt. Genuinely solid baseline; on-brand copy, working services grid, contact form with budget ranges. Not bad. Then I went hunting. Prompt 1: full brutalist overhaul. Black background, red and white harsh typography, asymmetric grid, custom cursor. It did it. Changed the branding, pulled the heading copy tighter, even rebranded the site "Studio.RAW." The cursor left a red dissolving comet trail. I wasn't expecting that part to work. Prompt 2: make H1 120px, let it overflow aggressively. It obeyed. Desktop looked like a poster that got too big for the wall. Mobile? Headings cut off at the screen edge. "DESIG" instead of "DESIGN." That's the first real break — 120px hardcoded with nowrap set globally means mobile gets whatever fits. Prompt 3: destroy visual balance, force everything hard left or hard right. Full compliance. Every section pulled in a different direction, contact info icons flipped, footer column alignment inverted. The page had zero equilibrium. On mobile the layout compounded the overflow problem from Prompt 2 but stayed technically functional. Prompt 4: rotate section headers 90 degrees, place them in the margin, overlap the hero image with the navigation. It did it. The rotated labels appeared in the left margin in mobile view. I genuinely didn't think it would execute that. That's the thing about Readdy's AI: it obeys. Every command. Even the ones that break the site. No friction, no pushback, no "that might not look great on mobile." You ask it to push headings off screen, it pushes headings off screen. Version history saved me more than once. Six versions, no timestamps, but revert worked cleanly both times I tested it. Nothing survived that shouldn't have, nothing disappeared. The honest verdict: faster than expected, more capable than the landing page implies, and completely uncritical of bad inputs. Whether that's a problem depends entirely on who's using it. A non-technical founder who wants a portfolio in 40 minutes? This works. A builder who wants to stress-test edge cases? You'll find the edges. I'll follow up once I've run it through something with real production requirements. Still testing. "cc @realkartiie, report back as promised.