The most useful meeting output was the list of decisions and owners.
Everything else faded.
Capture what needs to survive the meeting.
The rest is process.
A useful way to evaluate an AI model:
Don't just ask it something it knows.
Give it an unfamiliar problem, limited context and a chance to explain its reasoning.
That's where the interesting differences appear.
Reviewed last quarter’s bets.
Some worked, some did not.
The review itself improved the next set of bets.
Feedback on decisions matters as much as feedback on code.
Expansion happened because the product solved an adjacent problem without extra work from us.
Designing for natural extension points pays later.
Leave room for pull.