TLDR
I recently discovered that AI likes to use a standard template when creating landing pages. To get something unique you have to first break that concept by prompting your agent to be unstructured, random, unique and use compelling copy. Then I had it go through the below data loops in order to have it improve on its first version.
1) Does this webpage design feature stand out vs the majority of other webpages. If not update again
2) Is this webpage interactive and eye catching? If not update again.
3) Is this design element mobile friendly? In not update again
4) (campaign specific) Are elements which highlight my expertise most prominent on the page? If not update again
5) (campaign specific) Is all information in the copy accurate to my experience? If not update again
6) Would adding animation, color style or a unique graphic image improve this design element for the above criteria? If so update again
7) Is this cohesive with the rest of the website? If not update again
Every page on this site — the service pages, the portfolio, the one you’re reading now — was built the same way. I didn’t hand Claude a template and ask it to fill in the blanks. I gave it real numbers from real campaigns, told it to ignore the instinct to be safe, and then ran the result through the same eight questions before I’d ship it. That process is what I call data loop constants, and this article is the method itself, not a marketing description of it.
What Is A Data Loop Constant?
Look at the symptom picker on my services page, or the funnel diagram on the B2C page, or the holdout-test simulator on the CTV page. None of that content is a hand-written prompt into the page. Underneath each one is a small array — a constant — holding the real data: eight symptoms and which service fixes each one, five channels and what I actually buy and read on each, the real lift curve from a real flight-length test. The interface is a thin loop that reads the constant and renders it. Click a different symptom, the loop re-runs, the page updates.
That’s the whole trick. When I ask Claude to build something interactive, I don’t ask for “a tabs component.” I hand it the data first — the eight symptoms, the five channels, the real campaign numbers — and ask it to build the interface the data implies. The interactivity isn’t decoration bolted onto a template. It’s the data made visible, and a visitor clicking through it is actually learning something true about the work, not playing with a UI toy.
Why this matters for marketing, not just design
A generic template shows the same three bullet points on every service line. A data loop shows a different, correct answer for every input — which means every visitor who plays with it sees something tailored to what they clicked, built from real numbers instead of placeholder copy. That’s the difference between a page that describes expertise and a page that demonstrates it.
Why I Tell Claude to Ignore the Rulebook
Claude, like most AI writing and design tools, is tuned toward safe defaults — three-column grids, centered headlines, a hero image above a paragraph, bullet points in threes. Those defaults exist because they’re broadly correct and rarely result in offensive design outcomes. They’re also what every AI-assisted site now looks like, which means the safe choice has become the generic one.
So when I brief a page, I say explicitly that it does not need to follow conventional structure, format, copy length, or visual layout — that breaking from the expected pattern is the goal, not a risk. That one instruction changes what comes back. Without it, Claude reaches for the safest layout it knows. With it, Claude will propose a diagonal timeline instead of a card grid, a live simulator instead of a static stat block, sentence fragments instead of complete-sentence bullets — because I’ve removed the assumption that the conventional version is what I want.
What I AVOID Prompting and Prompt
Structure. I don’t ask for a hero, three feature columns, and a CTA. I ask what structure best represents this specific data — sometimes that’s a picker, sometimes a diagram, sometimes a single number that dominates the screen.
Format. Headings don’t have to be short and punchy. A few of mine run long, because the length of the sentence provides the right context, and a generic five-word header would have said less.
Copy. I tell it not to write like marketing copy — no “unlock your potential,” no adjective stacking. Flat, specific, sometimes blunt.
Design. Color, spacing, and motion don’t have to follow a predictable system. If a section is about urgency, the interface can feel urgent — tighter spacing, a harder color, a faster transition — rather than the same calm grid as every other section.
The risk is that “ignore the rules” can also produce something unusable. That’s what the eight questions below are for. Freedom on the front end, a hard filter on the back end.
The Eight-Question Loop
After Claude builds a design element, I run it through eight questions, in this order, and any “no” sends it back for another pass before I have it look at the next question. Here’s each one, and specifically why it earns its place in a marketing site rather than just a design checklist.
1. Does this stand out versus the majority of other webpages?
Goal: Would someone who has looked at twenty competitor sites today remember this page tomorrow?
Why it’s in the loop: Attention is the scarce resource in marketing, not information. A healthcare buyer comparing consultants has seen the three-bullet-and-a-headshot layout a hundred times; it registers as wallpaper. A page that looks different is the page that gets a second look, and a second look is the entire job of a homepage. This is the question that kills the default layout before it ships.
2. Is this design element interactive and eye-catching?
Goal: Does it entice a click, or does it just sit there being unnoticed?
Why it’s in the loop: Time on page is a trust signal, to both the visitor and to Google, and nobody spends ninety seconds reading static bullets. They spend ninety seconds clicking through five content areas to find the one that matches their problem. Every minute someone spends playing with the funnel diagram or the holdout simulator is a minute they’re absorbing the content — that I understand this problem well enough to model it — without feeling like they’re being pitched.
3. Is this design element mobile friendly?
Goal: At 390 pixels wide, does it still work — not just fit, but actually function with a thumb?
Why it’s in the loop: Most web visits happen on a phone, often from someone who isn’t at a desk — a family member searching mid-crisis, a marketer scrolling between meetings. A brilliant desktop interaction that breaks on mobile doesn’t just underperform, it damages the experience: broken layouts read as unprofessional in a category where trust is the entire sale. Everything ambitious I build gets tested at phone width before I call it done, not after.
4. Are the elements that highlight my expertise the most prominent thing on the page?
Goal: If someone reads only the biggest, boldest thing on the screen, do they walk away knowing I’m qualified?
Why it’s in the loop: This one is only applicable to a personal branding website. But the goal should be emphasized in the data loop. A slick animation that shows off Weight Watchers, Bradford Health and iHeartMedia in eight-point type at the bottom of the page has its priorities backwards. This question forces the hierarchy: the proof points the real numbers, the real names, have to out-rank the cleverness of the interface that’s displaying them.
5. Is everything in the copy accurate to my actual experience?
Goal: Could I defend every number and every claim on this element in a conversation, with specifics?
Why it’s in the loop: Again this is another goal loop and may not be applicable to most prompts. This is the one that isn’t really about design at all, and it’s the most important one on the list. An AI writing tool will smooth a real result into a rounder, more impressive-sounding one if you let it “over 100%” becomes “nearly triple,” a specific client becomes “a major health system.” In a category where a false claim is also a compliance problem, every figure on this site is a number I can trace back to a real campaign. This question is a hard stop, not a style preference: a beautiful design element built on an exaggerated claim gets rejected outright, no matter how well it scores on the other seven.
6. Would animation, color, or a unique graphic improve this against everything above?
Goal: Is there a specific addition that would raise the score on standout, interactivity, or prominence, not decoration for its own sake?
Why it’s in the loop: This is a deliberate second look after the element already “passes.” Good enough isn’t the bar. The CTV page’s holdout simulator didn’t start with a live chart redrawing as you drag the slider, that came from asking this question after the static version already worked. The addition has to serve one of the first five questions. Motion for the sake of motion gets cut in question eight.
7. Is this cohesive with the other pages on the site?
Goal: Dropped onto this page with no context, would someone recognize it as the same site they were just on?
Why it’s in the loop: Standing out from competitors and staying consistent with yourself are different goals, and it’s possible to win the first at the expense of the second. Every page here uses the same blue, gold and off-white system, the same typography, the same header and footer, but the interaction on each page is built from that page’s own data, so no two feel like the same template reused. Cohesive brand, non-repeating experience. This question is what stops “break the rules” from turning into a site that feels like six different freelancers built it.
8. Does this come across as a generic AI design?
Goal: The honest gut check, if I saw this on someone else’s site, would I think “that’s obviously AI-generated”?
Why it’s in the loop: This is the catch-all, asked last, after an element has already passed the other seven. Even a interactive, mobile-friendly, on-brand element can still have the tell; the purple-to-pink gradient, the rounded-corner card grid, the em-dash-heavy copy that reads like every other AI-assisted launch this year. If the honest answer is yes, it goes back, no matter how well it scored above. This is the question that’s only possible to ask well if you’ve actually looked at what “generic AI design” looks like. Which, at this point, is its own visual language, and worth learning to spot in your own work.s whether all that motion produces anything valuable. Writing a good verifier is the real craft of loop building.
What This Looks Like as an Actual Prompt
Concretely, here’s close to what I send Claude when I’m building a new interactive element. Not a template to copy verbatim — the data changes every time, but the shape of it stays the same:
Prompt structure I use
1. The real data first, as a plain list or table — not a description of the data, the actual numbers and names.
2. One line giving permission to break format: “This doesn’t need to follow a standard layout. Build whatever structure best represents this specific data, even if it’s unconventional.”
3. The constraint that keeps it honest: “Every claim has to trace back to a number in the data I gave you — don’t round up or generalize.”
4. After the first draft, the eight questions, one at a time, as follow-up prompts — not all at once, because each answer should inform the next revision rather than produce eight simultaneous edits that conflict with each other.
The order matters more than it looks like it should. Questions 1 and 2 push toward boldness. Questions 3, 4 and 5 pull back toward function, hierarchy and honesty. Question 6 pushes forward again, but only within what the first five already approved. Questions 7 and 8 are the final brand and taste check. Running them out of order — asking about mobile-friendliness before you’ve let the idea get ambitious — tends to produce something safe and forgettable, which defeats the first question before you’ve even asked it.
The Trade-Off Worth Naming
This is slower than a template. A page built this way takes longer to prompt, longer to review, and longer to fix when an element fails one of the eight questions and has to be rebuilt from the data up rather than restyled. I think that’s the right trade for a site whose entire pitch is “I do this work myself, carefully” — a site built off a template would be arguing against its own premise. If your business is volume rather than craft, this method is almost certainly the wrong tool; use a page builder and move on. If the page itself is meant to be the demonstration of how you work, it’s worth the extra passes.
If You Want to See It Rather Than Read About It
The B2C service page has the channel funnel this article describes in the opening section — click through Paid Social, Search, CTV, Influencer and App Store and watch the wire light up and the detail card rewrite itself. The CTV page has the holdout-test simulator from question six. The Services page has the symptom picker that opened this article. All three were built the same way: real data in, eight questions out, nothing shipped on the first pass.
If you’re building a marketing site with Claude, or thinking about it, and want to talk through how this applies to your own pages, that’s a conversation I’m glad to have — reach out here.