Training Coach

Reads measurements, progress photographs and workout logs, evaluates against the plan, and sends a check-in.

Reminding is cheap. Evaluating is the part that changes behaviour.

On the evening before a measurement day it reminds me. The next morning it pulls the numbers, the newest progress photographs and the workout export, compares all of it against the plan, writes the review into my notes and sends the check-in.

How it is built

Three inputs, three formats: hand-entered measurements, images from the photo library, and a CSV export from a training app. The AI agent reconciles them into one timeline.

Reading the photo library on a Mac is harder than it sounds, the application database is not a public interface, and the files you want are not where the interface suggests they are.

What it got wrong

It read a stale copy of the photo database. Copying the database file alone, without its write-ahead log, produces a snapshot that is missing the most recent photographs, which are precisely the ones a progress review needs. The AI agent reported “no new photos” on exactly the days there were new photos.

The lesson

A report-only AI agent still has to be right, because the failure mode is not a wrong action, it is a confident “nothing to report” that is actually “I could not see anything”. Those read identically from the outside, and only one of them is true.