Product AI
AI Agent
Built & Shipped
When an AI writes something confidently, that's exactly when it deserves a second look. Scoped disagrees with itself before you have to.
Role
Product Design & Full-Stack Build
INdustry
Product & AI Tools
CLIENT
Personal Project
OVERVIEW
It should feel like magic. That's the promise in every "use AI to 10x your workflow" reel — type a prompt, get a PRD, hand it to Lovable, done. I tried it once. It worked exactly as advertised, and that was the problem. So I built the tool that stops to ask questions instead.
I designed it, built it, and shipped it.
Try it live
Phase 1
The problem
Every AI PRD tool I looked at including my own first version - makes the same quiet assumption that a document that reads finished is finished. Nothing checks whether the stated goals agree with the stated scope. Nothing acknowledges the PRD is usually a dead end.. read once, then rewritten by hand into whatever prototyping tool or coding agent actually needs.
So I stopped trying to make it write better PRDs. I made it argue with the ones it already wrote.
Phase II
The Redesign
WHAT CHANGED
BEFORE
Grid lines, corner brackets, mono-spaced labels — read like an engineering spec, not a living document.
AFTER
Frosted glass and drifting color — a document that feels like it's still being formed, not measured.
SOLVING IT
1.
Every other tool stops the moment the PRD is generated. This one keeps going - it re-reads its own North Star and its own MVP scope and names where they contradict each other.
2.
The same PRD reads differently depending on who opens it. One click rewrites it three ways: engineering-first, a one-page executive summary, or a design brief leading with personas and user stories.
3.
A PRD only a human can act on is half-finished now. One button turns it into a CLAUDE.md-style spec for a coding agent. Another turns it into a tailored Lovable/v0 prompt, built from the PRD's actual features.
IN PRACTICE
Building a PRD generator is one kind of proof. Using it to build something is a harder kind. I ran a real idea through it — a group food-ordering feature — and took the generated PRD straight into Lovable with no editing.
01.
A tailored Lovable/v0 prompt, generated straight from the PRD's
own feature list and scope tensions.
No manual rewriting
Built from the PRD's actual constraints
BEFORE
Open Claude, write a prompt by hand, copy the output, hope it holds together, manually rewrite it for whoever reads it next.
AFTER
One input, a 13-section PRD, a built-in critique of its own logic, one click to reframe, one click to hand off.
REFLECTION
The hardest part of this build wasn't the generation, the deployment, or even the redesign - those were execution, even when execution meant fighting GitHub's folder structure for an hour. The hard part was deciding to let the tool undercut its own output on purpose.
It's easy to ship a feature that makes a tool look smarter. It's harder to ship one that makes the tool's own PRD look less finished, and trust that a willingness to disagree with itself is more useful than one that just sounds confident.
Confidence isn't the same as
correctness.