Coach

Coach is our own product. Our founder, Michael Wilkins, trains early: 3am sessions in a 24-hour gym in New Zealand, and at home when he is in London. The apps he had used log sets well, but none of them coach. He wanted one coach for training, food and sleep that adapts to where he is, the kit in front of him, how he feels that morning and how he is tracking against his goals.

He is not a developer and is not comfortable in a terminal. So the project doubled as a test of how we build software now: could a founder direct AI coding agents in plain English and end up with a real native app, not a prototype?

He set one rule for the design: design for 3am. Everything has to be readable and usable half asleep, on a Watch, between sets.

Claude Code, Anthropic's coding agent, wrote and tested all of it. A lead agent planned each piece of work and sent the large pieces to parallel agents, each working on its own copy of the code with its own simulators. Each piece came back with tests, and separate reviewer agents checked the code through five lenses at every major step, along with exercise form and usability. Michael never wrote a line. He made the product decisions, tested on his own devices and flagged problems from the gym floor with a Flag button built into the app.

Verification did the work that trust would otherwise have to do. About 1,120 unit tests run on a Mac in seconds, about 190 more cover the server gateway, and about 220 interface tests tap the iPhone and turn the Watch's Digital Crown in the simulator. Nothing reached his devices without passing, and both devices are backed up before every install.

Real use was the test plan. Each morning's session wrote a diary on the device, a log of screens, taps, Crown values and errors that never leaves it. The agents read the diary and fixed what it showed, often before he was home from the gym: starting weights that were far too light, a heart-rate gap on the treadmill, a dictation feature that took four seconds to start and returned nothing.

The same discipline applies to the AI inside the app. The coach can be spoken to at any point, mid-set included, but it cannot change anything directly. The model proposes changes as structured operations: swap an exercise, drop a set, rest today, move a session. The app checks each one against its own safety rules before applying it. Small changes apply at once with Undo, and bigger ones wait for a tap on Do it.