You can spot AI-built work instantly. So can everyone else. Here's the method that stops you getting the average: direction + references + correction + taste.
You can spot AI-built work instantly. The tells:
None of this is a model problem. It's an input problem. The model gives you the average of everything it's seen — because you asked it for the average. Generic in, generic out. Here's how to stop.
Don't ask AI to make something from scratch. That's how you get the average. You decide the direction; the AI executes it. Before you generate anything, answer: what is this exactly (one sentence)? Who's it for? What feeling should it give? What does it explicitly not look like?
Before you build anything, ask me 5 questions to understand what I actually want — the purpose, the audience, the feeling, the references, and what it should NOT look like. Don't generate until I've answered.
That one move kills 80% of generic output. You're directing, not delegating.
Drop your email and the full playbook opens right here — nothing to download.
The single biggest lever on quality. AI has no taste of its own — you lend it yours through references. Where I pull them:
The rule: don't reference anything that already looks AI. No purple gradients, no floating blobs. Pick work with a clear point of view. Save 5–10 strong references, then feed them in.
Here are 5 reference images [attach]. Don't copy them — extract the design language: the type treatment, the colour discipline, the spacing, the overall feeling. Then apply that language to what we're building.
The AI doesn't have the answers. It's guessing. A talented intern on day one — capable, fast, but it needs direction and correction. So never accept the first output. Correct it like you'd correct a junior:
Every correction is the actual instruction. The 3rd or 4th pass is where it starts looking like yours. The shortcut that changes everything: ask it to show you options, not describe them.
Mock up 3 versions of this on a single page so I can see them side by side. I'll pick one, then we'll refine it.
You can't judge a visual decision from a text description. Make it render. Choose with your eyes.
The hard one. AI can execute taste — it can't have it for you. If you can't tell good from average, no prompt will save you. The good news: taste is trainable. How I built mine:
Direction + references + correction + taste. Skip any one and you're back to the purple gradient.
The Front-End Design skill for Claude Code (Anthropic). The biggest single upgrade. It pushes Claude away from generic AI aesthetics — the purple gradients, the default fonts — toward distinctive, production-grade design with real typography, creative layouts, and thoughtful animation. Install it (one command in your terminal):
npx -y skills add anthropics/skills --skill frontend-design --agent claude-code
That drops it into your project's .claude/skills folder. Then just build as normal and tell Claude: "Use the front-end design skill to build this. Direction: [brutalist / editorial / premium SaaS / etc]." It outputs working HTML/CSS/JS or React with the design baked in. Pair it with steps 1–3 above and you're a long way past anything that looks AI-built.
Built this for Day 14 of Building With AI. If it was useful, the rest of the series is at @alixqureshi.