Monday morning, I sent the proposal. The client replied with four words.
Last Monday at 10 a.m., I sent a client, Zhang Xiao, an AI-generated homepage from a café. After two minutes, she said: “It looks like a template, not our brand.” I know the feeling: I used to ask AI to “make something more premium,” and every attempt looked like a pipeline product. I reworked it twice before figuring out the issue. It wasn’t that I couldn’t describe details; I hadn’t given AI enough real material to work with.
The method: feed it context before it starts designing
Anthropic designer Nate Parrott shared a practical idea in a video: instead of simply asking it to “design a page,” provide screenshots, sites you like, past work, and real pages from your existing product. Claude’s Artifacts and Claude Design can then be used for interactive prototypes, dashboards, and simple animations. The key is to have AI imitate a structure you have already validated, rather than inventing what “premium” means from scratch.
One scene stuck with me: in Nate’s San Francisco office, he turned an idea from his phone’s notes directly into a Duet feature sketch. He started by clarifying the goal, users, and existing interface, then revised it step by step. The useful part wasn’t a particular button. It was the loop: provide materials, make a draft, identify problems, adjust.
Replicating it today: modest budget, but set aside time for real communication
Here’s my estimate: roughly $20 per month to start, depending on the plan; around 40 minutes to prepare materials and revise the draft; the technical barrier is simply being able to type in a chat window and upload screenshots—no coding required. The first step isn’t opening a blank chat. I create a document and write three sentences: who this is for, what task it needs to handle, and which pages must stay. Then I upload three to five reference screenshots.
I got this wrong the first time too. I dumped in a dozen images with conflicting styles, which made the direction harder for AI to understand. Later I kept only three and labeled them “layout reference,” “brand color reference,” and “must-keep feature.” The page became much more stable immediately. I could also start with a narrow version rather than building the entire system at once.
How I’d approach it at different stages
If I were just starting, I’d practice with the free allowance and static pages by drafting one homepage or quote page for a single client. I’d use it to see whether they were willing to keep talking, without rushing into a long-term subscription. With one or two clients already, I’d organize their existing screenshots, brand colors, and previous deliverables into a “design reference pack” and reuse it before every generation, saving time on repeated explanations.
If I were scaling up, I’d have the design and product leads maintain separate references for core pages, available components, and prohibited styles, so nobody generates incompatible interfaces in parallel. We’d also test proposals against real tasks—for example, having a new colleague complete one workflow from the prototype—instead of only asking AI whether the result “looks good.” This tool isn’t necessary for everyone. If my business depends mainly on relationships, consulting, or in-person services, polishing one small deliverable may be a better investment than learning a design tool. There’s no urgency to try it either. I can wait until a client says, “It still doesn’t feel like our brand,” and come back with specific screenshots.