Let me be honest with you.
I didn’t save 3.5 hours. I compressed what used to take 3 weeks into 3.5 hours. That’s a different thing. And the difference matters if you actually want to use AI without becoming dependent on slop.
Here’s what I built: a complete brand identity, a design system, and a live website. From zero. For a real project.
Here’s what nobody tells you: 2 of those 3.5 hours were spent removing what the AI thought looked good.
The Setup
I opened Claude with a name, a rough concept, and a logo PNG.
No wireframes. No mood board. No 47-slide creative brief.
Just: here’s what this is, here’s who it’s for, here’s the mark. Build me a design system.
5 minutes later it handed me a full spec. Colors, typography hierarchy, spacing rules, button states, card styles, badge variants. A Design.md file that became the single source of truth for everything that came after.
That part? Genuinely impressive.
Then I looked closer.
The AI Slop Problem Nobody Talks About
Claude has a default aesthetic. It’s clean. It’s competent. It’s completely forgettable.
Left to its own judgment, it will give you:
• A hero section that looks like every SaaS launched in 2022
• Typography that’s “professional” the way stock photos are “authentic”
• Spacing that’s technically correct and visually dead
The output wasn’t wrong. It just had no point of view.
That’s the trap. Most designers using these tools ship the first pass and call it done. Then wonder why it looks like every other AI-generated brand on the internet.
I spent 2 hours making deliberate decisions about what to kill, what to push, and what to completely override. That’s not a bug in the workflow. That’s the workflow.
If you don’t have a design opinion walking in, the AI will fill the vacuum with mediocrity. Every time.
How It Actually Went Down
Step 1: Brand spec in regular chat. Not in the design tool.
This is the move that saved the most time. I used Claude chat, not the design surface, to build the full brand brief. Mission, audience, personality, color rationale, typography logic, voice and tone.
By the time I opened the design tool, I had a document. A real one. Not vibes. Not “make it modern.” A 300-line spec with opinions baked in.
The design tool executed. I directed.
Step 2: Design system first. Everything else second.
This is non-negotiable.
If you build the landing page before the design system, you’ll rebuild the landing page. Guaranteed.
The design system took about 25 minutes to get right. It saved probably 90 minutes downstream because every asset pulled from the same source. Same tokens. Same components. Same visual logic.
Step 3: Skip wireframes. Go straight to high fidelity.
I know what I want. I’ve been doing this long enough to have spatial intuition for layouts. Wireframes are for exploring. I wasn’t exploring. I was executing.
If you already know the answer, wireframes are just extra billable hours you pay to a machine.
Step 4: Image references over text prompts. Every time.
This is the one I want you to tattoo somewhere.
Text prompts produce approximations. Image references produce directions.
“Make it feel like a premium fintech product” produces 40 different interpretations.
Drop a screenshot of Linear’s dashboard next to a reference from a Stripe campaign and say “this tension, this density, this restraint” and you get something specific.
Every time I hit a prompt that wasn’t landing, I switched to a visual reference. Every time it worked faster, cleaner, and with less correction.
Reference-first is the actual unlock. Not better prompting. Not longer briefs. Just: show the machine what you mean.
The Website
I didn’t build it all at once. That’s the mistake everyone makes.
You prompt the whole page in one shot, Claude builds something that looks coherent on the surface, and you spend an hour untangling why it feels wrong everywhere simultaneously. You can’t fix what you can’t isolate.
So I broke it into pairs.
The whole page took about 45 minutes longer than doing it in one prompt would have. And it needed about 70% fewer corrections.
That’s the trade. Slower build, cleaner output, less cleanup. Every section was deliberate before the next one started.
What 3.5 Hours Actually Buys You
Not a finished product. A strong foundation that would’ve taken 3 weeks to reach manually.
The Part AI Can’t Do
AI will generate. It will not decide.
It won’t tell you that the hero feels emotionally flat even though it’s technically correct. It won’t know that the button radius is 2px too forgiving for the brand’s personality.
You have to know those things. And then you have to act on them.
The designers who are going to lose work to AI are the ones who were already deferring those decisions to clients or templates or trends. The tool just made that weakness visible faster.
The designers who are going to win are the ones who use AI to eliminate the mechanical parts of the job so they can spend more time on the actual design problem.
3.5 hours. One brand. One design system. One live website.
But the reason it doesn’t look like an AI project is because a designer ran the process.
That part doesn’t change.