On Monday morning Josh, Matt, and I got on a call to figure out the insights tab in Cico.
Cico is the calorie app the three of us have been building since July - calories in, calories out, the math that decides whether the week was a win. We knew we wanted AI in the insights tab to keep the content fresh. We did not, going into that meeting, know when a user should get a new insight, what kinds should exist, or where the tab should live.
An hour later we had answers to all three.
Granola was running the whole time. Josh and I have used it for months; this was one of Matt's first meetings on it. We weren't trying to take notes and participate at the same time. We were just talking.
I've been on both sides of the old PRD problem. At one extreme, a twenty-page Amazon-style document that exists partly to prove the person writing it has earned the right to ship the thing. At the other extreme, the early-stage startup that doesn't write PRDs at all because being busy feels like being productive. Both bury the idea. Both play telephone. Somebody had a thought, somebody else heard it, somebody else wrote it down, somebody else built from what got written. By the time the thing ships, it's not the same thought.
After the Cico meeting I opened our Cico Claude project and gave it the Granola transcript. The project has been getting smarter for months: every PRD we've written, our Cico design guidelines, access to the codebase. It knows what challenges are. It knows what the history tab does today. The skill expects a transcript as input and goes to work on two things: research, and questions.
The research is the easier half to explain. For the insights tab, it looked at how other apps surface AI-generated insights - not just nutrition, but finance and sports too. Pulled app store reviews. Scanned Twitter. The pitch isn't that Josh and I can't do this ourselves. It's that the research lives next to the spec by the time we read it.
The questions do something different. Meetings are incomplete; they have to be. A product gets discussed across weeks, not in a single hour, and an hour of recorded conversation contains a hundred quiet assumptions nobody surfaces because everyone thinks they're already shared. The skill pokes at those. You talked about insight frequency but didn't decide it. Daily, on-demand, or triggered by behavior? Half the time we'd already discussed the thing and forgotten to land it. Half the time we hadn't.
The history tab is the moment I knew this workflow had become the workflow.
Cico has a history tab. We'd been quietly debating, for weeks, what that tab is really for. Challenges already let you look at past challenges, so the history tab kept feeling thin. We'd parked it.
In the PRD it came back with for the insights tab, the skill proposed putting insights inside the history tab. The history tab was quantitative - a calendar of past days, the numbers. Insights made it qualitative too: the numbers, plus what they meant. Josh and I read it and looked at each other. We could have come up with that. We just hadn't.
The thing builders worry about with this kind of workflow is becoming less present in the meeting itself. If the AI is going to listen and synthesize, why pay full attention?
The opposite has been true. Josh and I are more focused in meetings now, because we know what we need to get out of them. The meeting has a job. The transcript has a job. The skill has a job. There's nowhere along the chain where the meeting itself can be lazy.
If you're a builder who keeps walking out of meetings annoyed that nothing got written down, the smallest version of this is almost embarrassingly simple. Get Granola, or whichever AI note-taker you like. Set up your project with the context it needs. Write or borrow a PRD skill that takes a transcript and gives you back a draft plus the questions you didn't answer. Paste in your next meeting and see what happens.
The meeting still has to be a good meeting. That part is on you. But the gap between we talked about it and we shipped it gets a lot smaller.