Claude Design: one prompt, three art directions, one shot

I gave Claude Design a real job: take an existing site and redesign it with full creative freedom, not a safe refresh. I selected Fable 5 and set reasoning effort to High, pasted one prompt, and made exactly one decision along the way. What came back was, honestly, close to a one-shot. Real content pulled from the live site, a clear point of view, the whole direction executed, and only minor tweaks left for me. The reusable part of that is the prompt, so that is what this post is really about.

The prompt

Here is the whole thing, verbatim. To use it, replace [SITE URL] with the site you want redesigned, select Fable 5, set reasoning effort to High, and paste it into Claude Design. That is the entire setup.

You are acting as a senior brand and web designer known for distinctive, award-caliber visual work, not a template assembler. Redesign this site with full creative freedom: [SITE URL]

Use the web capture tool and review the live site yourself, the homepage and enough interior pages to understand it fully. Do not ask me anything. Work this out on your own and proceed end to end in one pass.

Step 1, infer briefly before designing:
From what you find on the site, state in a short paragraph who the audience likely is, the top one to three actions the site wants visitors to take, the current brand personality, and what feels dated, generic, or weak about the design today. Also note what you believe must not change, such as core content, key data, or information architecture, versus what is safe to rework.

Step 2, choose a direction yourself, do not wait for approval:
Consider two to three distinct art directions internally, for example different moods, type systems, or color approaches. Pick the strongest one yourself. In a few sentences, name the directions you considered and explain why you picked the one you are building. Then execute it fully and immediately, without pausing for my input.

Creative brief:
Treat the current design as a starting point, not a constraint. You have permission to depart from the existing layout, color system, typography, and overall visual language. Do not produce a safe, incremental refresh, and avoid a generic SaaS template look. I want genuine visual ambition and a clear point of view.

Non-negotiables:
Preserve the site's actual content and information architecture unless a change is clearly better, and say so when you make one.
Meet WCAG AA color contrast and keep interactions keyboard navigable.
Design responsively and show how it holds up on mobile, not only desktop.
Use real content pulled from the site, not placeholder text.
Cover the homepage plus at least one representative interior page.

Why each part exists

The opening line is a role, not a compliment. Telling the model it is a senior brand designer known for distinctive work, explicitly not a template assembler, sets the bar the rest of the prompt keeps pointing back to. Without that anchor you tend to get a competent but forgettable refresh, the exact thing I did not want.

The instruction to use the web capture tool and review the live site matters more than it looks. I want the redesign grounded in what the site actually says, its real headings, its real data, not an invented version of it. Pairing that with “do not ask me anything” and “proceed end to end in one pass” is what turns this into a one-shot attempt instead of a back-and-forth. Every clarifying question the model might have asked is one I have preemptively told it to answer on its own.

Step 1 forces inference before pixels. Making the model state the audience, the top actions, the current personality, and what feels weak does two things: it surfaces its reasoning so I can sanity-check it, and it commits the model to a diagnosis before it starts designing against that diagnosis. The note about what must not change, core content, key data, information architecture, is the guardrail that keeps ambition from quietly deleting something load-bearing.

Step 2 is the part I care most about and, as it turned out, the part the model treated most loosely. The intent is explicit: consider a few directions internally, pick the strongest one yourself, and build it without waiting for me. The creative brief and the non-negotiables then fence the freedom. Depart from the current design, avoid the generic SaaS look, but hold the line on real content, WCAG AA contrast, keyboard navigability, responsive behavior, and coverage of at least one interior page. Ambition inside constraints, not ambition instead of them.

What actually happened

Here is the honest part. The model did Step 1 cleanly, a tight read of the audience and what was weak about the current design. Then, despite Step 2 telling it in plain language to pick the strongest direction itself and build immediately, it did not pick. It presented three fully fleshed-out art directions and waited for me to choose.

The three were genuinely distinct, not three shades of the same idea. “Sunrise Session” leaned into warm dawn light and a community-first feel. “The Project Board” was editorial and training-log in spirit, built around a “you are the project” framing. “Night Circuit” went for big poster energy, dark with electric accents. I picked Night Circuit.

I should flag the deviation plainly: the prompt says decide yourself, and the model asked me instead. But I will also say it was a happy deviation. Being handed three real directions, each thought through rather than hinted at, and choosing with conviction beat having a single pick imposed on me. If the model had silently committed to Sunrise Session I would have wondered what I was missing. Seeing all three and landing on Night Circuit felt like a better decision, not a worse one. I would not change the prompt to prevent this. I would just tell you to expect it.

The one shot

Once I validated Night Circuit, the rest really was close to a single shot. The model executed the direction fully rather than sketching it: real content pulled from the live site, the homepage and an interior page, the dark-plus-electric language carried through consistently, and the non-negotiables respected. What was left for me was minor, the kind of small tuning you would do on any handoff, not a rebuild. For one prompt and one decision, that is a lot of finished work.

The follow-up: selling the redesign

The part I did not expect is what happened next. In the same conversation, right after the redesign was approved, I ran a second prompt, and this is where it got interesting. Instead of starting from scratch, the second prompt deliberately builds on everything already sitting on the canvas: the mockup, the new design system, the homepage, the interior page, the mobile views. It does not describe those screens to the model, it tells the model to go use the ones it just made.

The job this time was a client-facing sales deck, the kind of thing you would put in front of a non-designer decision maker who never saw the process and only cares about the result. Same setup as before: replace [CLIENT NAME]with whoever you are pitching, keep Fable 5 on High, and paste. The result was, again, close to a one-shot. A deck that pulled the actual screens into before-and-after comparisons and dressed itself in the redesign's own visual language.

You are acting as a senior brand strategist and presentation designer, not a slide-template filler. In this project, review the website redesign mockup already on the canvas, the new design system, the homepage, the interior page, and the mobile views. Build a client-facing sales presentation deck that sells this redesign to [CLIENT NAME], a non-designer decision maker who did not see the process, only the result.

Do not ask me anything. Work this out on your own in one pass.

Ground the deck in the real work:
Pull actual screens from the mockup, homepage, interior page, and mobile views, and place them directly in the deck. Do not redraw or approximate them. Take the deck's own visual identity, colors, type, spacing, and mood, from the mockup's new design system, so the presentation itself demonstrates the new direction rather than looking like a generic slide deck wrapped around some screenshots.

Narrative, use this as guidance, not a rigid outline, and adapt slide count, aim for roughly 8 to 12 slides:
Title slide naming the project.
Where the current site falls short today, two to three concrete pain points, stated factually, not as criticism of the client.
The design vision in one clear idea, the single sentence that captures the new direction.
Before and after for each key page, current screen next to the new one.
Mobile, shown as its own moment, not an afterthought.
What this changes for the business, clarity, credibility, conversion, accessibility.
What stays intact, content and structure that did not need to change.
Next steps.

Design bar for the slides themselves:
No generic template look. Strong typography, generous whitespace, one idea per slide. Minimal text, since this will be presented live and read from a distance, not read as a document. Every slide should look intentional, not assembled from a stock layout.

Storytelling tone:
Persuasive and confident, but honest. Do not oversell or manufacture problems that are not real. Let the before and after comparisons carry the argument.

When the deck is ready, prepare it so it is exportable and presentable to the client, PDF or PPTX. Do not offer anything beyond the finished deck.

Why it works comes down to the same moves as the first prompt, pointed at a new target. It leads with a role, a senior brand strategist rather than a slide-template filler, so the model aims for something intentional instead of stock. It grounds the deck in the real screens already on the canvas, telling the model to place the actual mockups and explicitly not to redraw or approximate them, which is only possible because the first prompt left that work sitting right there to harvest. It has the deck take its own visual identity from the new design system, so the presentation demonstrates the direction instead of wrapping a stock template around some screenshots. It gives a narrative as guidance rather than a rigid outline, a suggested arc the model can adapt. And it asks for honest persuasion, letting the before-and-after comparisons carry the argument rather than manufacturing problems. One prompt made the work; the next one sold it, off the same canvas.

The honest caveats

This is one test, not a benchmark. I ran it once, on one site, with one model and one effort setting, and I am reporting a subjective impression of a single strong result. I did not run it ten times to see how often the model actually self-selects a direction versus offering three, and my one data point suggests it may not always follow that instruction to the letter. A different site, or a different run, could go another way.

With that said, the takeaway I trust is about the prompt. Give the model a strong role, make it ground itself in the real site, force a diagnosis before design, hand it bounded creative freedom, and hold firm non-negotiables. That structure is what produced a redesign worth keeping on the first serious attempt. Swap in your own URL, keep Fable 5 on High, and see what it does with your site.