Giving AI freedom over content — not the design
Designing a sales-page generation system where AI writes structured course copy while designer-built templates retain control of layout, hierarchy, responsive behavior and dynamic LMS components.
Learn product design by shipping, not by watching.
Twelve weeks where every module ends in a critique with a working designer. Built for people who already have a job and no patience for theory.
The Friday critiques were uncomfortable in the way that actually changes how you work.
Aarav ShahI came in able to make screens and left able to explain why they were those screens.
Meera IyerSome product visuals have been simplified or reconstructed to protect confidential platform and customer information. Course names, copy and data are fictional.
Every course could be different. Its sales page wasn’t.
The existing experience used the same fixed structure across very different education businesses, so three unrelated offers arrived at the same page with the title swapped.
Customization existed. Self-service did not.
An educator could change almost anything — which was exactly the problem. Choosing a template was the easy part; everything after it was writing, structuring, configuring and validating a website.
The research question wasn’t “what does a nice landing page look like?”
It was what a course buyer needs before they commit — which turns the page from a layout into a sequence. Thinkific, Leadpages and a set of course sales pages were studied to interrogate that sequence, not to copy it.
- 01AttentionHeadline
- 02RelevanceProblem and outcome
- 03TrustInstructor and testimonials
- 04UnderstandingCurriculum
- 05DecisionPricing and call to action
- 06ObjectionsFAQ
Automation could remove the blank page. It could also remove the design.
The template should own the conversion logic. AI should own the words inside it.
Some decisions are made once, by a designer, for every page. Others are made per page, from the course in front of you. Sorting them is the whole design.
The same page, three generations.
Regenerate and watch what moves. The copy changes completely; hero position, CTA placement, sales card, syllabus, pricing, reviews and responsive behaviour do not.
Learn product design by shipping, not by watching.
Twelve weeks where every module ends in a critique with a working designer. Built for people who already have a job and no patience for theory.
The Friday critiques were uncomfortable in the way that actually changes how you work.
Aarav ShahI came in able to make screens and left able to explain why they were those screens.
Meera IyerGeneration was deliberately constrained.
AI replaces text. Classes, styles, hrefs, images and layout are not touched — which is what makes the output safe to ship rather than safe to demo.
Headline, subheadline, description, story, benefit copy, FAQ and CTA copy. Text nodes, and only text nodes.
Markup, CSS and classes, layout, hierarchy, links, images, responsive behaviour and widget contracts.
Designers create the safe space AI works inside.
A template is not a picture of a page. It is a validated structure — theme tokens, dynamic widgets, responsive behaviour and quality checks — tagged to the segments it suits, and re-ingestable after edits.
A template can be reusable without being generic.
Generic categories rarely match the reference an educator holds for their own business. The school segment narrows the starting pool before AI writes a word — reuse that gets more specific, not less.
Generate from where the work already happens.
No separate AI product. Generation starts in the builder the admin is already in, and the publish decision stays theirs.
Learn product design by shipping, not by watching.
Twelve weeks where every module ends in a critique with a working designer. Built for people who already have a job and no patience for theory.
The Friday critiques were uncomfortable in the way that actually changes how you work.
Aarav ShahI came in able to make screens and left able to explain why they were those screens.
Meera IyerThe AI does not only write. It can also notice what is missing.
Suggestions name the missing ingredient; they don’t silently insert it.
AI variability becomes layout variability.
A headline can arrive at two words or twelve, and the template has to hold either. Content tolerance became a design requirement, not a bug report.
AI entered an existing system.
The builder already had templates, widgets, a theme system, publishing and responsive behaviour — all validated before a template is available to generate against. This project replaced none of it and added one thing: structured text.
Four contracts carried the boundary into the build.
Each one was something I could check afterwards in design QA.
People used the generated starting point.
Generation quality is not measured by how much text AI writes. Three signals tell you whether a generative feature is actually working — and only one of them is about the model.
A product target set for the feature. It has not been reached.
- Time to sales page — did generation remove work?
- Publish rate per generation — was the output useful enough to ship?
- Average edits per page — how much correction did the first output need?
What this changed in how I design generative products.
AI needs a clear boundary of authority.
Templates turn generative variability into scalable quality.
The first useful output matters more than the most flexible one.
Existing structured data should stay deterministic.
What I’d measure next
- Measure generation → publish by school segment
- Understand why admins edit generated pages
- Compare conversion between generated and legacy sales pages
- Test content-length resilience across languages and course types
This project taught me that AI needs a clear boundary of authority.