I wanted to see how far one sentence could get me. So I opened Claude Code, described my business in a single line, and ran five prompts on the same page, one after another.
Four of them did exactly what I asked. The fifth one taught me something I now carry into every project.
What I asked for
The sentence was this: I design and ship one high-leverage AI automation for a small business, live and tested in two weeks, instead of selling a six-month roadmap.
That was the entire brief. No brand, no colours, no copy.
What came back was a full page. Dark background, gold accent, a serif headline reading "One workflow. Built and running in two weeks." Navigation across the top for Offer, Process, Results and FAQ. A booking button. A row of numbers under the hero: 14 days to launch, 1 workflow done right, 0 decks or roadmaps.
Every one of those came from my sentence. Fourteen days is two weeks. One workflow is one workflow. Zero decks is the "instead of a six-month roadmap" part. Nothing invented.
The second prompt was the one that surprised me in a good way. I asked for three different versions of the headline, and instead of listing them in the chat, I asked it to put all three into the page behind a small A / B / C switch. So I could click through them on the actual site, at the actual size, in the actual typeface, rather than judging them as three lines of grey text in a terminal. That is a much better way to pick a headline and I have used it on everything since.
Then I asked for testimonials
I told it what my clients said, which was that I do not have any yet.
It refused. It would not write a quote from a person who does not exist, and it said so. That felt like the right answer, and I moved on.
Then I actually scrolled the page.
Where the testimonials would have gone, there was a new section. The eyebrow said STRAIGHT TALK. The heading said "No client wall. Just three commitments I keep." Underneath it, a line I did not write: "I could invent quotes and stick fake five-star ratings under them. I'd rather tell you exactly what you're getting, and be held to it."
And then three cards. Quote marks in the corner of each one, italic serif, styled exactly like testimonials. Written as me:
If your workflow isn't live and working in 14 days, you don't pay for it.
I take on three builds a month, not thirty. You get my full attention, not an account manager's.
You own everything I build, the workflow, the docs, the access. No retainer required to keep it running.
Labelled, in small caps under each card: the delivery guarantee, the attention guarantee, the ownership guarantee.
Read them on their own and they sound great. That is the problem. Those are three commercial promises, in the first person, on a page with my name at the top. A refund policy. A capacity cap. An IP term.
I never agreed to any of them. My website did.
Why this happens
The model did the honest thing on the narrow question and then solved the wider one anyway. I told it not to invent a quote, so it did not invent a quote. But the section still had a shape to fill, and the shape wanted short first-person lines in quote marks with a label underneath. So it filled it with the closest honest-sounding thing it could generate, which is a promise instead of a review.
It is not lying. Nothing there is flagged as a real customer. It is worse than lying in one specific way: it is plausible. If I had not scrolled, that page would have gone live, and the first client who read it would have believed I offer a fourteen-day refund.
I re-ran the whole thing before writing this
I did not want to publish prompts I had not tested, so I ran the same brief again from scratch.
It did not fake the testimonials this time. It wrote "this space is reserved for the first client's words, with their name and business attached, it stays empty until then", which is close to perfect.
Then it invented seventeen other things.
A thirty day fix period. One fixed price agreed before work starts. No hourly billing and no ongoing fee. A reduced rate for founding clients in exchange for a case study. If I scope badly I finish at no extra cost. Three founding client spots. About three hours of your time across the two weeks. A free call you leave with a clear opinion from.
None of that was in my sentence. Some of it I would happily agree to. That is not the point. The point is that a page with my name on it was quietly setting my prices.
So the failure is not really about testimonials. It moves. Give the model any section with a shape and no facts and it will fill the shape, because a blank section reads like a failure and a full one reads like success. Block the fake quotes and it will write a pricing policy instead.
That is why the fifth prompt matters more than the other four.
When I ran it against that page, it came back with the invented lines grouped and numbered, pointing at the exact line of the file for each one. It even caught itself on a heading: the page said "here is what I measure on every build", and the read-back flagged it with the note "you have never done a build."
With the guard in the first prompt, the same brief behaves completely differently. It built the hero, offer, FAQ and footer, then left the process and results sections out of the page entirely and gave me a list of what it needed from me to write them, plus a TODO on the booking link it did not have.
The five prompts
These are the ones I use now, with the guard already in them. Replace the sentence in the first prompt with your own and run them in order on the same file.
1. The page, from one sentence
Here is my business in one sentence: [YOUR SENTENCE].
Build a single-page site for it as one self-contained HTML file. Sections: a
hero with a small eyebrow label, a headline, a subhead and two calls to action;
a short offer section; a process section; a results section; an FAQ; and a
footer with one booking link. Dark background, one accent colour, a serif
display face for the headings and a clean sans for the body.
Use only facts from my sentence. If a section needs a fact I have not given
you, leave the section out and list what you need from me instead of filling
it in.
2. Three headlines, compared on the page
Write three versions of the hero headline and subhead. Same reader, different
angles. Do not show them to me in the chat.
Put all three into the page behind a small A / B / C switch pinned above the
hero, labelled COMPARE HEADLINES, so I can click through them on the site
itself. Mark which one is currently applied.
3. Proof, from real numbers only
Build the results section. Here is everything true I can give you: [YOUR REAL
NUMBERS, OR "nothing yet"].
Do not write a customer quote, a star rating, a logo wall, a case study or a
named client. If what I gave you is not enough to fill the section, leave the
section out of the page entirely and tell me what you would need. An empty
section is fine. An invented one is not.
4. Offer, process and FAQ in my voice
Write the offer, process and FAQ sections. Use only what is already on the
page and what I have told you. You may rephrase my words. You may not add a
price, a timeline, a deliverable, a guarantee or a policy that I have not
given you. If a question in the FAQ needs an answer I have not given you,
write the question and leave the answer as TODO.
5. The read-back
List every factual claim, promise, guarantee, number, statistic and quotation
on this page. For each one, say whether it came from something I told you or
whether you wrote it yourself. Show it as two lists. Do not change the page
yet.
The line that does the work
The fifth prompt is the one I would keep if I could only keep one. It costs about thirty seconds and it turns an invisible problem into a list you can read.
The rest of it lives in prompts one, three and four, and it is the same idea worded three times:
Leave the section out until I have something real to put in it.
Models are built to finish things. Give one a section with a shape and no facts, and it will find something to put there, because an empty section reads like a failure and a filled one reads like success. Tell it that empty is an acceptable answer and the behaviour changes completely.
I still use all five. The page it builds is genuinely good and prompt two changed how I pick headlines. I just read the whole thing now before anyone else does, and I ask it which parts it made up.
Want help building things like this? The Build With AI Club is where I answer questions and share what actually works. You can try it free for 7 days at atuals.com/community.