Nord LuceCase Study
One person, an AI team, and a live AI product.
Thinking about putting AI to work in your business? This is how I built a full AI product with AI agents doing the building and me directing them. You'll see how the work was split between me and the AI, what went wrong, and the rules I now bring to every AI automation I build for a client.
The Problem
Ask an AI to write for you and it comes back sounding like every other AI. You can describe your style in a prompt, but most people can't say what makes their writing theirs. They can only tell when something sounds right or wrong.
I had the same problem and couldn't find a tool that let me show an AI how I like to write, so I built one: LLM Voice Lab.
What I Built
Voice Lab turns a writer's reactions into rules. A writer chooses between pairs of writing samples, the engine works out the rules behind their choices across thirteen dimensions of style, and it exports a portable voice file they can hand to any AI.
Six dimensions are free. The extended set unlocks for 12 US dollars. It runs at llmvoicelab.com on Cloudflare, Render and Supabase, with Stripe for payments.
How It Was Built
I directed the work and made the product decisions. A team of AI agents did the building: one agent wrote each piece, a different agent reviewed it, and a final check ran across the whole build.
Every call to an AI model got a written contract for what information it needs, with checks in the code that stop the run if something is missing, so the engine fails loudly instead of quietly guessing.
What Went Wrong
The agents built faster than anyone checked whether the product did its one job. Tests passed and reviews came back clean while the writing it produced still missed. Passing tests showed the code held together. They couldn't show that the product captured a voice. That took a person reading the output.
The product was first called ready in mid-August. The version I deployed came 45 days later, and part of that time went to rework that a person reading the output earlier would have caught.
What I Do Differently Now
One thin path from start to finish comes first, on the real system, before any feature gets deep. Each piece is finished only when a person has used it and read what it produced, not when the tests pass. And nothing goes into the build that the current step doesn't need.
These are the rules I bring to every AI automation I build for a client.
Where It Stands
Voice Lab is deployed and in an invite-only beta. Payments run in test mode until the beta closes. I'll add real usage here once there is some to report.
Book a Call
Thinking about building something with AI? Tell me what it should do. I'll tell you how I'd build it and where it could go wrong. No charge, no obligation.
CamCamrick Schoen-Clark, Nord Luce