1/ I curated 7,842 lines of markdown to fight AI slop. Not a product. Not a framework. A **skill file** for AI agents that build websites. Because every AI agent I work with produces the same five patterns: Purple-to-blue gradients. Centered hero. Three identical feature cards. "Trusted by 10,000+." Lorem ipsum in disguise. That's not design. That's the default output of an LLM. ๐Ÿงต --- 2/ The patterns are predictable because they're in every training set ten thousand times. `linear-gradient(135deg, #667eea 0%, #764ba2 100%)` is the visual shorthand for "an AI made this." `border-radius: 9999px` on every button is the structural shorthand for the same thing. If your output looks like this, it doesn't matter how good your prompt was. It reads as generated. --- 3/ So I wrote a skill that says "no." Not in a hand-wavy way. With **28 specific anti-patterns**, each with an example and a replacement. Not "avoid generic design." Instead: *Purple-to-blue gradients are slop. Replace with warm paper + ink + editorial red.* Not "use good typography." Instead: *Fraunces + Inter + JetBrains Mono. Hero at 60โ€“160px. Display tracking -0.035em. All-caps tracking +0.14em.* --- 4/ It's modular. 18 files. Each loadable independently. ``` SKILL.md Identity, principles, process aesthetics.md 7 style directions *-patterns.md 22 sub-styles (Linear, NYT Mag, Bandcamp, ...) typography.md Type system color.md Token system layout.md Grids, spacing, responsive anti-patterns.md What to reject components.md What to build motion.md What to animate content.md What to write a11y + perf The floors most agents skip checklist.md Pre-ship QA ``` The agent pulls only what it needs. Doesn't burn context on irrelevant guidance. --- 5/ The most important file is `aesthetics.md`. It defines 7 directions โ€” Refined Minimal, Editorial, Swiss, Brutalist, Soft, Technical, Playful โ€” and tells the agent to **pick one, commit to it, don't mix them.** Because "modern web design" as a single style is itself AI slop. The best sites are opinionated. The skill teaches the agent to be opinionated. --- 6/ It also teaches the agent to **write specific copy.** โŒ "Empowering businesses to thrive" โœ… "Ship features 3x faster" โŒ "Welcome to [Brand]" โœ… "Design that doesn't need explaining." โŒ "Fast. Simple. Beautiful." โœ… "A magazine for readers, not scrollers." The cardinal rule: write the way you'd talk to a smart friend, not a marketing department. --- 7/ Last thing: a pre-ship checklist of 70+ items. Not "does it look good?" โ€” that's vibes. Specific questions: - Hero headline 60โ€“160px? - One accent color, used <10% of pixels? - No `transition: all`? - Focus-visible defined on every interactive element? - Real names, real numbers, no lorem ipsum? - `prefers-reduced-motion` honored? - Would a senior designer ship this? If 6+ answers are "no" โ€” keep iterating. --- 8/ It's open source. MIT license. If you build AI agents that touch frontend โ€” Cursor, Claude Code, Cline, custom โ€” this should be in your context window. If you design and use AI as a tool, this is the missing piece between "AI-generated" and "AI-assisted." Link in next tweet. /fin --- 9/ [link to repo / gist] If it works for you, ship something good with it. That's the thanks. If it doesn't โ€” open an issue. The skill is meant to evolve with the slop it pushes against.