Notes
Field notes from building it ourselves
What worked, what failed quietly, and what we'd do again, from building Routine Run and the Structed AI interview. Every number here was measured on a real system.
Apple put a language model in every new iPhone. Here's what it's good for
What Apple's on-device model did well in Routine Run, where it failed quietly, and the test that keeps a prompt and its parser from drifting apart.
Let AI assistants into your product without handing them the keys
How Routine Run's MCP connector lets Claude read runs and act, while every change a model writes waits for a person to approve it on the phone.
Our AI interview was losing facts clients gave it. 18 of 19 came back
Three fixes from the Structed AI interview: a safety net that matched the wrong thing, a question asked until the turn limit, and a trace lookup that made replies take a minute.
The translation tax: when half the work is understanding what the client meant
Marina Shmayger found her £60 hourly rate was really £20. The missing hours went to translating what clients meant. That's the problem an AI interview takes on.
Writing it down doesn't mean it has left your head
Why most processes fail at automation: they were written down but never run. Do it, understand it, refine it, and only then automate the steps that earned it.
Record it, watch it back, rebuild: the loop behind a pitch and an AI agent
At the Grishin Robotics accelerator, founders pitched, got recorded and watched themselves back every week. The same loop is how you improve an AI agent.