In plain words: if you build a website with an AI agent and don't specify a palette, you get the same cream-and-serif page as everyone else. I'm not a designer and I can't invent a palette, so I stopped trying to fix the look and fixed the explaining instead — I drew my customer as a comic strip, with people in every frame. The page still has the cream background. It no longer looks like everybody's.
The "AI aesthetic" is a real, documented thing, not my imagination: cream and beige backgrounds, rusty orange accents, big italic serif headlines, wide-letterspaced subheads. Kyle Chayka catalogued it in The New Yorker, and the reason is statistical rather than aesthetic — a model returns the median of what it has seen, and the median of recent web design is a small set of defaults. Telling the agent "make it different" mostly swaps one fixed palette for another.
I hit exactly this. My page for teams was correct, organised, and completely flat. Someone read it and said: it looks like a Medium article on a cream background. That was accurate, and my first instinct — pick a braver palette — was the wrong instinct, because I would have been guessing at something I have no skill in.
The reframe that helped: a landing page exists to make a product understandable, and the cream is not what makes it hard to understand. Words are. A page of claims asks a stranger to agree; a story lets them recognise. And recognition gets much stronger when the story contains drawn humans, because a reader looks at a figure at a desk and sees themselves at their desk — no argument required.
So I kept the palette, which is honest for a site about people working with agents, and put people into it. Six scenes about one customer: what's on his screen, who's in his doorway, what he doesn't know is happening two floors away. My rule was that every scene had to be drawable, and that rule did real work — anything I couldn't draw turned out to be an abstraction I hadn't thought through.
The drawing was engineering more than art. I ran one identical prompt through three image models before making any real panel, and picked the one that produced clean schematic line art rather than the prettiest. Then the trick that solves character consistency: accept one panel, and pass that image as a reference into every later generation. My characters kept their faces and line weight because panel one was the anchor, and each person got exactly one identifier — round glasses, curly hair, a striped sweater — so a reader can follow them without a single name on the page.
Four rounds of criticism followed, and every note was about story logic, never drawing quality. The executive in the doorway was twice human size and read as caricature; authority is about 20% taller, not 100%. The "meanwhile, two floors away" panel didn't land until both desks had the same monitor, so the reader connects them instead of being told. And the panel about the colleague who quietly helps everyone failed twice — until the applause in the second half went to one of the people he'd helped. That version needed no caption at all.
Same argument, same facts, same cream background. Completely different page — and the words underneath now explain what the pictures already made you feel.
Reproduce this
Prerequisites: a page that reads as correct and flat; somewhere to serve WebP; and an image model you can call repeatedly with an image reference — that reference capability is the hard requirement, because it's what keeps characters consistent.
I used Magnific, which is a paid service — the models I used require a premium plan. Budget honestly: the line-art model I settled on costs 15 credits per image, the two I compared it against cost 60 each, so my 36 renders came to roughly 630 credits. Expect to spend 4–6× your final panel count, because rejects are part of the method.
Cheaper paths that satisfy the same requirement: a local Stable Diffusion or Flux setup with an image-reference workflow (free after setup, slower to get consistent), or any hosted image API that accepts reference images. Do not pick a model that can't take a reference — you'll spend the savings on characters whose faces change every panel.
Your agent will need from you: who the customer is, and — the part only a human supplies — which drawn scene is a lie. You'll know instantly. The agent won't.
Steps:
- Stop trying to out-design the default palette unless you actually have that skill. Decide whether the default is dishonest for your product; if it isn't, leave it and spend the effort on comprehension instead.
- Write the customer's day as 5–7 scenes, not sections: what they see, what they feel, what they don't know. Refuse any scene you can't draw — the refusal is the useful part.
- Put a human in almost every frame. That's the whole mechanism: readers recognise a person at a desk faster than they parse a benefit block.
- Before generating anything real, run one identical prompt through 2–3 models and compare. Choose for legibility at thumbnail size.
- Generate the first panel only and iterate on proportions. Subtle exaggeration reads as real; heavy exaggeration reads as cartoon and loses the recognition you're buying.
- Pass the accepted panel as an image reference to every remaining panel, and repeat each character's single identifier in every prompt.
- Save every variant as its own file and review them in a folder, not in a chat. I kept 36 renders for 7 used panels; the rejects teach you what the model does with your style.
- Write captions as HTML under the image, never inside it — crisp, translatable, and rewritable in seconds once you can see the pictures.
- Write the page copy last, as explanation of what the strip already shows.
- For two audiences, swap only the opening panel and the caption that bridges into panel two. One strip, two entry points, no duplicated page.
Verify: show the page to someone who has never seen your product and ask them to describe the customer's problem back to you using only the pictures and captions — no body copy. If they can, the strip is doing the explaining. Then check every caption: it should say what changed, not name what's visible. A caption that describes its own picture is noise.