Over the past few months, I've been building ToolMorph, a collection of developer tools and AI experiments. One idea I was particularly excited about was Founder Roast, an AI-powered tool that reviews startup landing pages and gives actionable feedback instead of generic suggestions.
When I started building an AI-powered landing page reviewer, I thought the hardest part would be writing a good prompt.
I was wrong.
The real challenge wasn't AI at all—it was everything around it.
Over the past few months, I've been building ToolMorph, a collection of developer tools and AI experiments. One idea I was particularly excited about was an AI that could analyze a startup's landing page and provide brutally honest feedback.
Not generic advice like:
"Improve your CTA."
I wanted feedback that sounded like it came from an experienced founder, designer, marketer, and developer reviewing the page together.
Something that answered questions like:
- Would I understand the product in five seconds?
- Does the value proposition make sense?
- Is the page trustworthy?
- Would I actually sign up?
It sounded simple.
It wasn't.
The First Problem: AI Can't Review a Website It Can't See
Large language models don't browse websites the way humans do.
Giving the AI a URL like:
https://example.com
doesn't magically let it understand the page.
That meant I first had to collect the page myself.
I built a pipeline that:
- Downloads the HTML.
- Extracts meaningful content.
- Removes navigation, scripts, and unnecessary elements.
- Preserves headings and structure.
- Sends only the useful information to the AI.
This alone turned out to be a much bigger engineering problem than I expected.
Too Much Context Is Just as Bad as Too Little
My first implementation sent almost the entire page to the model.
The results were... terrible.
The AI would spend paragraphs discussing footer links, cookie notices, or navigation menus while completely missing the product itself.
So I changed the approach.
Instead of sending everything, I focused on the parts a real visitor notices first:
- Hero section
- Headline
- Subheadline
- Primary CTA
- Key features
- Social proof
- Pricing (when available)
The quality of the responses improved dramatically.
Sometimes giving an AI less information produces much better answers.
Prompt Engineering Wasn't About Making the AI Smarter
Initially my prompts looked something like this:
Analyze this landing page.
The responses were polite, generic, and mostly useless.
After several iterations, I realized the model needed a role, a process, and constraints.
Instead of asking for opinions, I asked it to evaluate specific categories, explain why something worked or didn't, and provide actionable improvements.
The difference was night and day.
The AI stopped sounding like a chatbot and started sounding like someone performing a real product review.
Every Website Is Different
One startup might have:
- beautiful design
- weak messaging
Another might have:
- excellent copy
- confusing navigation
Others hide their pricing, overload visitors with animations, or bury the call-to-action halfway down the page.
Building a system that could produce useful feedback across all of these cases was much harder than optimizing for a single layout.
There wasn't one perfect prompt.
There were dozens of edge cases.
Good AI Products Aren't Just Prompts
One lesson surprised me the most.
The prompt wasn't the product.
The product was everything around it.
- Collecting reliable page data.
- Cleaning noisy HTML.
- Choosing the right context.
- Handling failures gracefully.
- Structuring the AI's output.
- Presenting feedback in a way people could actually use.
The model was only one piece of a much larger system.
Try It Yourself
If you're curious about how your landing page comes across to a first-time visitor, I built Founder Roast as part of ToolMorph.
It analyzes your landing page and highlights things like:
- Whether your value proposition is clear
- If your call-to-action stands out
- Trust and credibility signals
- Messaging that might confuse visitors
- Practical suggestions to improve conversions
You can try it here:
👉 Founder Roast: https://toolmorph.in/labs/founder-roast
Or explore the rest of the developer tools and AI experiments on ToolMorph:
👉 ToolMorph: https://toolmorph.in
What I Learned
Building this project changed how I think about AI products.
I used to believe success came from finding the perfect prompt.
Now I think prompts are only a small part of the equation.
The real value comes from building systems that feed models the right information, ask the right questions, and present the answers in a useful way.
That's the part users actually experience.
What's Next?
I'm continuing to improve the reviewer by making its feedback more specific, more actionable, and better at understanding different types of startup websites.
If you're building AI products, I'd love to know:
What's one engineering challenge that turned out to be much harder than you expected?
Top comments (0)