I built an AI tool that turns a plain-language goal into a professional-grade prompt. 86 people hit the page. 93% left, and 2 finished. The tool was not broken. The design was: 8 questions, an email gate at question two, and a name only a prompt engineer could love. The rebuild asks 2 questions, lets the AI do the structuring, and delivers the prompt before asking for anything. Completions went from 2 out of 86 to roughly 2 in 3. This is the whole log, with the numbers.
Most people get garbage out of AI because they put garbage in. I kept watching smart business owners type one vague sentence into ChatGPT, get a mediocre answer, and conclude AI is overrated. So I built a tool to fix the input side: the Prompt Architect. It walked you through the same structure prompt engineers use, and at the end you had a genuinely great prompt.
Then I watched the data, and the data said almost nobody made it to the end.
What the funnel actually showed
These numbers come straight from the tool's own analytics, and they are the whole reason this post exists.
Two completions out of 86 visitors. And the landing page promised "three simple steps" while the tool delivered a wall of text. When the page says simple and the product says homework, people believe the product.
Why do users abandon AI tools?
In our case, friction before value. The Prompt Architect opened with your first name, then your email. Before you had seen anything the tool could do, it was already collecting. Then came the real work: your role, your objective, the task, the context and background, your constraints, and the output format. Eight questions total, every one of them a place to give up.
Each field was legitimately valuable. That was the trap. I was thinking like someone who knows AI, uses it professionally, and enjoys the craft. The person I built it for does not know what "role" means in a prompt. They do not know the role they need. They only know the outcome they want.
"Valuable steps make a valuable tool." Reality: a tool made by a prompt engineer for people who are not prompt engineers is backwards. I even named it the Prompt Architect. The name alone told normal people this was not for them.
The rebuild: two questions instead of eight
The fix was not adding features. It was subtracting friction. The rebuild, renamed the Prompt Sherpa, works like this: tell it what you want in plain language, it asks you one smart follow-up question, and it hands you the finished prompt. Two questions. The email ask moved to after the prompt is delivered, once the tool has already proven itself.
Every piece of the old structure still exists. The role, the objective, the task, the context, the constraints, the output format: all still in the final prompt. The user just never has to author them. The AI reads the goal, asks its one question, and builds the structure on the back end. Same output quality, none of the homework.
In the demo run, I asked it to create Facebook ads for a social media marketing company. It asked one follow-up: who is the audience and what is the offer. Thirty seconds later I had a prompt with the full structure, including constraints like no scare tactics and no healthcare policy violations, plus instructions telling the AI to work with me back and forth instead of dumping a wall of output. Pasted into a fresh chat, it set up a working environment for building ad campaigns, and the first thing the AI did was ask clarifying questions. Exactly as designed.
And the funnel flipped. Since the rebuild shipped, the tool's analytics show 60 to 68% of visitors finishing and walking away with their prompt. The 8-question version converted 2%. Same tool underneath, 6 fewer questions in front of it.
What a good prompt contains (and who should build it)
- A role: the expert the AI should act as. Users do not know this. The AI can infer it from the goal.
- An objective and the specific task. Users know this part. It is the one thing they should type.
- Context and background. One smart follow-up question collects what matters.
- Constraints: policies, tone, what to avoid. Inferred from the use case, not requested from the user.
- Output format, and how the AI should collaborate. Baked in by default.
The structure was never the problem. Asking the user to supply it was.
The rules this rebuild runs on
You do not need to be technical to build with AI. Both versions of this tool came from Lovable and Claude Code, and I am not a developer. But you do need to be willing to ship something ugly, watch it fail, and believe the data over your ego. 93% of people told me my product sucked. They were right, and the numbers do not care how clever the form design felt.
If you ship it and it is pretty, you might have shipped too late. Garbage in, garbage out still runs everything. And when a product is dying, look for what to remove before you look for what to add.
Frequently asked questions
Why do users abandon AI tools?
In our live data, friction before value. Our prompt tool asked 8 questions, with email as question two, before giving anything back. Of 86 people who hit the page, 93% left and 2 finished. Cutting the form to 2 questions and moving the work to the AI took completion to 60-68%.
What makes a good AI prompt?
A role, an objective, the task, context, constraints, and an output format. Users should not have to author those pieces themselves. An AI can build the full structure from a plain-language goal plus one smart follow-up question.
How many questions should an onboarding form ask?
As few as it takes to deliver the first win. Our 8-question version converted 2 of 86 visitors. The 2-question rebuild (state your goal, answer one follow-up) made the same tool usable, and we ask for the email after the value is delivered, not before.
Do you need to be a developer to build an AI tool?
No. Both versions of this tool were built by a non-developer using Lovable and Claude Code. Building it was never the hard part. Designing it for people who do not live in AI all day was.
The point of all this
This is what an AI department looks like.
Real tools shipped fast, real funnel data read honestly, and a rebuild driven by what users actually did instead of what felt clever. We run our own products this way, and we install the same system inside client businesses.
Book a 30-minute fit call