Claude does the work. You decide what good looks like.
When anyone can ship, care, intention and taste are the differentiator. Claude doesn’t have taste. You do.
Based on my talks at the Claude meetup in Melbourne on 5 November 2026 and at Outlook: AI on 12 November 2026.

I’m super lazy. I want to use AI for everything. But I’ve been coding since I was 12 and doing graphic design since I was a teenager, and I don’t want to lower the bar to get there. Be lazy, but still have high standards.
Ask Claude for a homepage and leave it alone, and you get Claudisms: purple glowing borders, glass, an “AI-powered” badge, a pulsing “Live” dot, uppercase monospace. It’s horrendous. The copy is plausible and way too verbose. It doesn’t know who it’s talking to, so it panders to everyone and doesn’t really say anything. Any one of these is a choice. Three on a page and it looks generated.
Claude will go rogue and add too much crap. Your job is to dial it back, reel it in, and work out what you’re actually trying to say. Relentlessly remove elements. Simplicity above all else. Don’t make the user think.
Nobody ships AI slop on purpose. They just can’t see it. That’s why I built Taste Training, which is almost like Duolingo for visuals. The first drill shows two signup cards and asks which one a person actually decided on. It’s the one that says what the product does. The other could be anyone’s.
Four headlines
Where I start is I look at the best examples out there and copy them precisely: Linear, Things, Raycast, Mintlify. They’re clean and to the point. One headline, one line, one button. Copy them, refine the edges, work backwards to why they’re good, and you start to understand what good looks like. Then hand Claude the reference.
Generating is cheap now, so on Done Bear, my to-do app, we tested four homepage headlines. The original, “One list. Every device.”, is still live.
I once asked Claude to optimise that homepage for answer engines like ChatGPT and Perplexity. It over-indexed on the SEO results and stuffed the page with a timestamp down to the minute, a key facts block and FAQs. It made no sense. You have to use your own judgement to dial it in.
Voice drifts the same way. The Done Bear voice guide has two columns: what we say, and plausible drift. “A powerful workspace for modern productivity” is what you get when nobody’s steering. Your brand book is a prompt.
Thirty-nine words to twelve
I ran the same brief twice on the same model, taking the first run each time. Claude’s default subheadline was 39 words, every feature in one breath. For the second run I wrote the prompt by hand: pages I like, our voice guide, and a checklist to run before it showed me anything. It came back with 12: “It works on Mac, Windows, Linux, iPhone and the web, even offline.”
That prompt took real thought, so I wanted everything I know about graphic design and software development in repeatable patterns that anyone can use, without being a developer. I write them as agent skills: a checklist Claude reads before it starts, with each rule written as the problem in plain English, then the fix. They’re open source and have been installed more than 33,000 times.
The typography skill has 78 rules, and Claude breaks a few of them every day: straight quotes, a hyphen where a dash belongs, an x for times. They’re tiny, but your eye reads them as care. The UI skill names 19 slop signals, and the fix is usually to take the thing away. The writing one, Ghostwriter, strips out the Claude and Codex isms and learns how you write from your own WhatsApp, Slack and email. I did a blind test and couldn’t tell what I wrote and what the agent wrote. Okay, the system works.
An emoji in my own example
If all these rules live in skill files, why not have the AI enforce them every time you make a change? Taste Lint does that: design review on every commit. When I ran it on my talk slides, the finding that blocked the build was an emoji in my own example.
The judgement calls, like a vague error message or an empty state with no next step, don’t fit a fixed check. Those go to Jev, from TypeSafe AI, which answers with a probability instead of a yes or no.
AI loves to say it’s not X, it’s Y. Cut it. Any failure from Claude that rubs you the wrong way, put it in as a rule, and you’ll never see it again.
Every time a new model comes out, I do a pass through all my skills. Sometimes I can chuck stuff out because the models have got smarter.