Four projects he builds daily with Claude Design

The first example highlights his core methodology. Lacking formal animation training, he first had Claude Design build a custom video editor, then used that editor to create his animation. His core principle: have AI build a custom tool first, then use it to execute your work.

The remaining three examples follow the same pattern: generating interactive widget controls with sliders, then using those tools to fine-tune hard-to-describe parameters.

10 practical tips from the creator of Claude Design

Here are the actionable takeaways. While the original post listed ten items linearly, we restructured them into three logical phases reflecting your actual workflow: pre-design preparation, active drafting, and final polishing. Each phase addresses a specific inefficiency.

1. Think it through before sending prompts

This runs counter to common habit. Most people sit at their screens and improvise, relying on AI speed for cheap trial and error. Nate does the opposite: he spends substantial time refining prompts, doing most of his thinking away from his computer.

Whatever medium works best for you, the key is doing the thinking away from your computer. That way, when you sit down, Claude can execute a clear, fully formed idea.

Nate Parrott

2. Turn repetitive patterns into a design system

Upload your brand files and visual assets, and Claude will analyze them to build a tailored design systemA standardized repository of brand colors, typography, buttons, and UI components ensuring consistency across projects.. From that point on, every project begins with your established design rules rather than a blank canvas.

3. Feed it your actual codebase and context

If you are designing for an existing app or website, connect it to GitHub. Claude will ingest your components and existing screens as a baseline, matching your real app's look and feel after a few prompts. It can also browse the web or connect to internal systems for live data.

4. Point at the screen and speak

Instead of writing a lengthy paragraph describing a specific UI element, select it directly on the canvas and speak your feedback.

5. Lock in fonts and colors upfront—don't let it decide

Without explicit direction, Claude defaults to its preferred internal aesthetic preferences. This is the single most practical tip, as the cost of omitting style guidelines is remarkably specific.

What exactly is this default aesthetic? A June feature in The New Yorker detailed its signature traits:

The article highlighted a telling example: independent designer Matt Ström-Awn recalled two separate clients presenting AI-generated sales decks. Both featured identical structures—a light background with a mission statement and three bullet points on slide one, a four-card grid on slide two, and a centered "our move" tagline on slide three. "They looked like they came from the exact same agency—different logos, identical design," he noted.

Anthropic acknowledges this tendency in its user documentation:

[The model] has strong design instincts that incline toward a consistent house style... that default is always there.

Anthropic Documentation, cited via The New Yorker

Here lies the critical trap: Anthropic's docs caution that negative prompts like "don't use cream" typically nudge the model into another static preset palette rather than fostering genuine diversity. General negative instructions fail; you must provide explicit font and color specifications—tell it what to do, not just what to avoid.

6. Request ten versions, then remix

Nate's tactic is asking for ten variations in one go. Most will miss the mark, but one or two will hit. You then prompt: "I like Option B, but mix in elements from Option D. Give me five hybrids blending both."

7. Sketch it when words fail

When you have a layout in mind but struggle to articulate it, draw a quick rough sketch on paper and snap a photo. Simple boxes, arrows, and labels communicate far better than three paragraphs of text.

8. Start with wireframes when details don't matter

Requesting a wireframeA low-fidelity structural blueprint establishing element placement without colors or custom fonts. is much faster and offers a major key benefit: it forces Claude to focus on structural layout rather than visual styling details.

9. Polish the last mile yourself—don't prompt for micro-tweaks

For final adjustments—rearranging elements, deleting text, resizing, or tweaking colors—use direct editing drag-and-drop tools rather than sending new prompts to Claude.

Prompting "shift this slightly right" requires the model to estimate what "slightly" means, leading to iterative guessing. Dragging the element directly gives instant visual confirmation in a single stroke.

Connecting Tip 6 and Tip 9 reveals a vital strategy: requesting ten versions consumes significant tokens, but doing so helps find high-level direction. Spend tokens on discovering direction, not pushing pixels. Once direction is set, minor tweaks like shifting positions or font sizes should be handled manually.

10. Queue prompts instead of waiting

You don't need to wait for Claude to finish generating before sending new instructions.ueue multiple prompts or type follow-ups while it works; Claude will process them sequentially.

The complete 10-tip checklist

Everything you draw should be alive

Nate concludes by recommending a famous 2012 talk by Bret Victor. Known for eschewing slides in favor of live, interactive demos, Victor inspired an entire generation of digital product designers.

The talk, titled Stop Drawing Dead Fish, stems from a sharp observation: artists draw fish as static figures frozen on paper. But living fish swim, dodge, and react to their environment. While computers possess the inherent ability to bring drawn objects to life, we continue using them to produce dead drawings.

People are living things that act and react. Creations inside a computer can also be living things that act and react... if we permit them to be.

[...] Everything we draw should, by default, be alive.

Bret Victor · Stop Drawing Dead Fish (2012)

Revisiting the beginning of Nate's article through this lens reveals the central theme running through his work.

What was the root cause of his initial bottleneck? Static design mockups are dead. Drawing screen after screen, arranging every state, and manually stringing them together merely simulates interaction. Engineers, meanwhile, ship living software that clicks, animates, and responds. The velocity gap between design and engineering was ultimately a gap between static artifacts and living systems.

His solution was not to "draw dead fish faster." He stopped drawing static mockups altogether to build living prototypes directly. HTML is the ideal medium because it is inherently alive: interactive, responsive, animated, and easily shareable via a single link.

His favorite creations share this quality: interactive documents, conversational slide decks, hybrid chart-video experiences, and designs that serve as their own editors. They defy conventional categories because traditional categories were built for static media.

Viewed through this lens, the ten tips fall into place: defining goals, locking in styles, choosing between variations, and polishing the final details represent human judgment that AI cannot replace. AI handles execution and scale. You define what the creation should be; the machine brings it to life.

The three phases address distinct inefficiencies: preparation stops vague prompting; drafting avoids single-direction focus; polishing prevents wasting tokens on minor pixel tweaks. Tip 2 (Design System) organizes logos, fonts, and colors into reusable components, so every project starts with your brand guidelines rather than a blank canvas.

This tip carries the clearest financial and design consequences. Left unguided, Claude falls back on a distinct default aesthetic detailed in Kyle Chayka's June 24, 2026 article in The New Yorker [Editor's Note].

The critical insight: Anthropic's documentation acknowledges this default and warns that negative prompts like "don't use cream" merely push the model into another static preset palette. Prompts must specify exact fonts and colors.

He is explicit about two core boundaries.