AI landing page prompt: the bad one, the good one and why
Guide/ 12 min read
Why AI keeps giving you the same landing page, and how four inputs fix it: audience, goal, brand and what to avoid. With a full prompt to copy.
- AI
- Shopify
- CRO
In this article (7)
A good AI landing page prompt does not ask for "a nice landing page". It gives the model four things it cannot guess: the audience, the goal, the brand and what to avoid. Leave them out and the AI fills each gap with the most common choice, so you get the same page as everyone else.
Below we put the bad prompt next to the good one, line by line, using a Shopify store in Mexico as the example, and then ask for a second and a third pass without starting over. Every recommendation comes from Anthropic's prompting best practices guide. If the page is already code and something breaks, the next read is how to report a bug to AI.
Why AI gives you the same landing page every time
This is the prompt nearly everyone writes the first time:
Make me a landing page for my coffee store.
- "Make me a landing page" names the format but not the job: selling one product, collecting emails, announcing a launch?
- "for my coffee store" gives the category and nothing else. It does not say who buys, where they come from, how the brand talks or what you are tired of seeing.
Whatever the prompt leaves out, the model decides for you, and it picks the most likely option. Anthropic's guide says it plainly in its frontend design section: without guidance, models can default to generic patterns that create what users call the "AI slop" aesthetic. Its sample prompt lists the symptoms: overused font families (Inter, Roboto, Arial, system fonts), clichéd color schemes, "particularly purple gradients on white backgrounds", predictable layouts and cookie-cutter design that lacks context-specific character.
Scroll the figure sideways to see all of it
The same guide offers a test for clarity: show your prompt to a colleague with minimal context on the task and ask them to follow it. If they would be confused, Claude will be too. It frames Claude as a brilliant but new employee who lacks context on your norms and workflows: the more precisely you explain what you want, the better the result. A designer on their first day who got "make me a landing page for my coffee store" would come back with questions before opening their tools. If you do not answer those questions for the model, it answers them itself.
The 4 inputs that change the result: audience, goal, brand and what to avoid
You do not need a huge prompt. You need one that contains what the model cannot know on its own. For a product landing page, that fits in four inputs:
| Input | What to write | If it is missing, the model assumes |
|---|---|---|
| Audience | Who buys, where they arrive from (ad, search, email) and on which device | A generic visitor, on desktop, who already knows the brand |
| Goal | One primary action and the exact text of its button | Several calls to action competing with each other |
| Brand | The concrete fact that sets it apart, the voice, hex colors with their use, and the typefaces | Default colors and fonts, interchangeable copy |
| What to avoid | The generic patterns you do not want and, for each one, what to do instead | The usual: centered hero, three cards, filler reviews |
Scroll the figure sideways to see all of it
Two recommendations from the guide make those inputs work harder:
- Give the reason, not just the rule. The guide recommends explaining why an instruction matters, because Claude can generalize from the explanation. Its example: instead of "NEVER use ellipses", say that a text-to-speech engine will read the response and will not know how to pronounce them. On a landing page, "do not ask for an email before checkout: this traffic comes to buy or to leave" tells the model what to prioritize in every section, not only in the form.
- Say what to do instead of what not to do. It is one of the guide's tips for controlling the format of a response. That is why, in the example, everything to avoid comes with its replacement: not just "no generic headlines" but "instead, one concrete fact".
One note on the avoid block: do not use it to describe your brand in the negative ("we don't want anything boring"). Use it to name the specific patterns you have already seen on other pages and do not want repeated.
A complete AI landing page prompt for a Shopify store in Mexico
The example brand, Bruma Alta, is fictional: a coffee roaster from Coatepec, Veracruz, selling online across Mexico. All of its data (ages, cities, grams, colors) is illustrative; swap in your own. The prompt is in English, but it asks for the page copy in Mexican Spanish, because that is what the store's customers read. You can paste it into a Claude chat or hand it to Claude Code.
You are going to propose a product landing page for a Shopify store in
Mexico. Context first; at the end, what I want you to deliver.
<audience>
People aged 28 to 45 in Mexico City, Guadalajara and Monterrey who
already drink specialty coffee but buy it at coffee shops. They arrive
from an Instagram ad, on their phone, and have rarely heard of us.
</audience>
<goal>
Get them to add the 340 g bag of medium roast to the cart. One primary
action only: the "Agregar al carrito" button. Do not ask for an email
before checkout: this traffic comes to buy or to leave.
</goal>
<brand>
Bruma Alta roasts coffee from Coatepec, Veracruz, every Tuesday, and
every bag has its roast date printed on it.
Voice: warm and concrete, informal "tú", no superlatives.
Colors: forest green #1F3B2D and cream #F3EBDD. Terracotta #C4622D only
for the button, so it is the one thing that stands out.
Type: a serif for headlines and a plain sans for body text.
</brand>
<voice_examples>
<example>Tostado el martes, en tu puerta el viernes.</example>
<example>Chocolate y mandarina, sin amargor de quemado.</example>
<example>¿Prensa francesa o V60? Elige tu molienda.</example>
</voice_examples>
<avoid>
- Generic headlines like "Welcome to our store". Instead, one concrete
fact: origin, roast date or tasting notes.
- Purple gradients and rows of identical cards with icons. Instead, the
product photo and one short list.
- Made-up reviews, numbers or awards. Instead, leave [REAL DATA] where
one is needed and I will fill it in.
</avoid>
<deliver>
1. The page sections, in order, with one line on what each is for.
Treat each one as a section of the Shopify theme.
2. The copy for every section, in Mexican Spanish.
3. What shows on a phone before scrolling, and where the button sits.
</deliver>
Here is how it reads, block by block:
- The first two lines say what the task is, for which platform, and in what order the information comes. That is the guide's "be clear and direct" part: the model knows from the start what it is reading.
<audience>describes who the page talks to and how they arrive: from an Instagram ad, on a phone, without knowing the brand. That drives almost everything else: short copy, the product up top, nothing that assumes they know you.<goal>sets a single action and the exact button text, and explains why there is no email gate. That "why" is the context the guide tells you to add.<brand>opens with the fact that sets the store apart (the roast date on every bag), then gives the voice, the hex colors with what each one is for, and the type. The terracotta comes with its reason: the button should be the only thing that stands out.<voice_examples>holds three lines, each in its own<example>tag. The guide calls examples one of the most reliable ways to steer format, tone and structure, recommends 3 to 5 of them, diverse and wrapped in tags so Claude can tell them apart from instructions. They are in Spanish because they are the brand's real voice, and each has a different shape (a delivery promise, tasting notes, a question) so the model does not lock onto one pattern.<avoid>names three specific patterns and, for each, what to do instead. The third one shuts the door on invented reviews or numbers: a gap marked[REAL DATA]beats a fake quote on your store.<deliver>is a numbered list of what you want back. The guide recommends numbered steps when order or completeness matters, and being specific about the output format. Asking for each block "as a section of the Shopify theme" gives you a structure that maps cleanly onto the theme editor.
Why tags like <audience>? The guide explains that XML tags help Claude parse prompts that mix instructions, context and examples without ambiguity, and asks for consistent, descriptive tag names. There are no magic names: <audience> works the same as <publico> as long as you stick to one. They also let you point at each block in later passes.
If you work in Claude Code, the guide's frontend design section links to the definition of the frontend-design skill, which pushes in the same direction as the <avoid> block. Installing that one and others is covered in Anthropic's official skills. A skill does not replace the four inputs: it has no idea who your customer is or what you sell.
How to iterate: the second and third pass
The first answer is a draft. The guide describes the most common self-correction pattern as: generate a draft, have Claude review it against criteria, then have Claude refine it based on the review. For a landing page that means two more passes, each with a different ask.
Scroll the figure sideways to see all of it
Second pass: review against your own inputs
Before changing anything, check your proposal against <audience>,
<goal> and <avoid>. Tell me which parts do not work for someone arriving
from Instagram on a phone, and why. Then rewrite only those parts.
- "Before changing anything" splits the review from the rewrite. You see the diagnosis first and can push back before any copy changes.
- "against <audience>, <goal> and <avoid>" uses the first prompt's blocks as the criteria. You do not invent new ones; you check against what you already asked for.
- "someone arriving from Instagram on a phone" restates the audience's concrete situation, which is the first thing to fade in a long draft.
- "rewrite only those parts" keeps the model from redoing the whole page and losing what already worked.
Third pass: specific, numbered changes
Three changes and nothing else:
1. Headline: use the roast date instead of "specialty coffee".
2. Move the "Agregar al carrito" button so it shows without scrolling
on a phone 390 px wide.
3. Remove the "Nuestra historia" section and keep one sentence about
Coatepec below the price.
Leave everything else as it is.
- "Three changes and nothing else" sets the limit before the list.
- Each number is a change you can check: the headline does or does not show the roast date, the button is or is not visible without scrolling at 390 px wide, the section exists or it does not.
- "Leave everything else as it is" protects what you approved in the second pass.
A rule of thumb, ours rather than the guide's: if the first answer gets the structure wrong (three calls to action, a newsletter block above the product), do not fix it with more passes. Go back to the prompt, fix the input that was missing or badly written, and start a new conversation. Passes polish a page; they should not turn it into a different one.
Common mistakes
- Adjectives instead of inputs. "Modern, clean and attractive" tells the model nothing it would not do anyway. "Terracotta #C4622D only for the button" does.
- Several goals at once. Selling, collecting emails and telling the brand story on the same page gives you three competing calls to action.
- Banning without an alternative. "No gradients" leaves the gap open; "instead, the product photo" fills it.
- Leaving out the device. If your traffic is mobile and you do not say so, nothing in the prompt asks the model to design for a narrow screen first.
- Leaving room for invented data. A model fills gaps; without a marker like
[REAL DATA], it may write a testimonial that never happened. - Asking for everything in the second pass. Change the headline, colors, order and tone at once and you cannot tell which change made the page better.
Frequently asked questions
Does this prompt work with other AI tools, not just Claude?
The recommendations here come from Anthropic's guide, which is written for Claude. The four inputs do not depend on the tool: whatever you leave out, a model fills with the common choice. What can vary is how each model responds to tags and formatting.
Do I have to use XML tags?
No, but the guide recommends them when a prompt mixes instructions, context and examples, as this one does. Use descriptive names and keep them consistent, which also lets you cite them in the second pass.
How long should a landing page prompt be?
Long enough that a colleague with no context could build the page without asking you anything. The example fits on one screen. If yours grows much longer, check whether you are describing the whole page instead of giving inputs.
What if my brand has no style guide?
The minimum is enough: your logo colors in hex, two or three lines you have already published and the one concrete fact that sets you apart. That helps the model more than a long brand book nobody pastes in full.
Checklist
- The prompt says who buys, where they arrive from and on which device.
- There is one primary action, with the exact button text.
- The brand has a concrete fact, a voice and hex colors with their use.
- There are three to five voice examples, different from each other, each in its own tag.
- Everything to avoid comes with what to do instead.
- Reviews, numbers and awards stay as
[REAL DATA]until you have your own. - The second pass reviews against your own blocks before rewriting.
- The third pass asks for numbered changes and leaves the rest alone.
A prompt with context does not guarantee a good landing page, but without it the AI has nowhere to get your brand from. If you then turn the page into code with an agent, install Anthropic's official skills, follow my 4-step Claude Code workflow, and when something does not work, use the right way to report a bug to AI. To decide what to test on the page next, continue with CRO on Shopify without guessing.
Saw this on Instagram? Comment PROMPT on the Reel and I'll send you the guide.
Written by

Published October 30, 2026 · updated November 1, 2026