I write the functional specs that development, QA and the Product team build on, and Training then carries them to the whole company. They took me 20 hours each and I had become the bottleneck. This is how I rebuilt the process, iteration after iteration, until specs started writing and updating themselves.
Writing a spec used to take 20 hours. Now it takes 3, and it updates itself after every meeting.
Every feature needs a functional spec before development. I write three to five a month, and each one took about 20 hours.
Developers, QA and the Product team all depend on that document, and through Training it reaches the whole company. While I was writing, they were waiting: the spec had become the bottleneck that set the pace of the roadmap. And by the time it was finished, the next meeting had already changed something.
I didn’t want to cut corners on quality. I wanted to stop spending my time on the parts that didn’t need me.
I broke the 20 hours down to see where they went. Most of it wasn’t thinking about the product: it was drafting, copying screens from Figma and chasing decisions that had been made out loud in meetings. Each of those had its own fix.
Writing a spec took about 20 hours.
The AI drafts each section with its own expert agent: business goals, acceptance criteria and more.
Decisions were made in meetings and never reached the spec.
Meeting recordings of ideation and solution talks feed the spec, and follow-ups update it.
Screens and details were copied from Figma by hand.
Connected to Figma, the AI reads the design and pulls the screenshots itself.
I didn’t automate it in one go, and it took far more than five tries. I spent three weeks iterating only on the spec structure and the AI skill that writes it, spec after spec, until the output was something I’d sign. Then I automated one bottleneck at a time.
Throughout: the AI creates the spec in Google Drive and keeps updating that same document.
Specs are more complete and up to date, so development asks less during the sprint.
Decisions from meetings reach the spec, so fewer things get built differently from what was agreed.
The hours I got back went into talking to clinics and deciding what to build next.
Manual first. A great manual spec was the blueprint for the AI. If I hadn’t known what a good spec looked like, I couldn’t have taught it.
Iterate the process. I didn’t design the final system up front. Three weeks went into the structure and the AI skill alone, and each step removed the bottleneck the last one left.
Decisions stay human. AI removes the busywork, but the spec is only as good as the decisions behind it. My job moved from writing to deciding.