Back to home

Artificial intelligence Live 2026

WebSpec AI

From a vague idea to a specification a team can execute.

Project creation screen in WebSpec AI: a form with the project name and idea, and flow templates (Full Spec, Landing Page, MVP SaaS, Desktop App, API Backend).

Where it came from

Working with AI on my own products confirmed something I already knew from client projects: the bottleneck was never writing code. It's ambiguity. An agent that gets a vague instruction produces a lot of wrong code, faster than ever.

The problem

Between "I have an app idea" and "this can be built" there is work almost nobody does: brief, PRD, RFC, UX definition and backlog. Without it, the project starts anyway, but decisions get made one at a time, late and without context.

The solution

A conversational application that guides the user through that sequence of documents, in order, building each artifact on top of the previous one until it reaches an exportable specification. Instead of jumping straight to code, it produces the product package that makes the code the easy part.

What's solved

  • Guided flow: brief, PRD, RFC, UX/UI, backlog, specs and final export.
  • Each document builds on the previous ones instead of being generated in isolation.
  • Conversational interface: the user answers questions, they don't fill in forms.
  • Public website presenting the product.

Evidence

  • Product website published and accessible.
  • Application deployable to Netlify with configuration included.
  • No secrets in the repository: the API key is set through an environment variable.

The hard parts

  • A flow that's too rigid feels like a form, and one that's too loose never reaches a specification. Finding that middle ground was nearly all of the product work.
  • The documents have to be consistent with each other: if the PRD changes, the backlog that came out earlier is now lying.

What I learned

  • It's the same method I apply to my own products —specify before building— turned into a tool.
  • Output quality depends far more on the design of the question flow than on the model behind it.

Next steps

  • Direct export to formats that coding agents can consume.

Applied learnings

Reading that grew out of this project