Skip to content
localwork
Go back

Twenty-Seven Fewer Sets

I’ve logged my strength training in GymBook for years: every exercise, set, rep and weight. What I wanted to know was what all those days add up to. Am I actually getting stronger? How is my training changing?

Asking Claude or ChatGPT meant exporting my data and uploading files by hand, every time. So I started building GymBook Intel, a small local app that tries to answer those questions from GymBook’s data. I’m building it iteratively with AI coding agents. This covers its first three days in Git, 8 to 10 November 2026.

GymBook stays the source of truth

I log in GymBook and nowhere else. GymBook Intel only reads. I didn’t want to switch trackers or move years of history by hand.

Getting the data across had to be close to one action. An iOS Shortcut takes GymBook’s database export from the share sheet and saves it to iCloud Drive as GymBook-latest.database, overwriting the previous one. On the Mac, a Dock icon runs a refresh and opens the dashboard. The refresh never modifies that file. If a new export has fewer records than the data in use, it stops. Exporting is still a step I take myself.

What the database actually says

GymBook stores its data in SQLite. Before drawing anything, I needed to know what the records meant: which sets were warm-ups or skipped, how weight was stored. And GymBook has no session entity. A set belongs to an exercise, a day and a routine, but there’s no record of a workout as such.

One extraction step turns GymBook’s tables into a single plain one. Everything is Python and SQLite from the standard library. Five reps and ten reps at different loads aren’t the same performance, so only performances with the same reps at a single load are compared. I left out an estimated one-rep max on purpose. It would look precise while resting on assumptions the records can’t check.

Charts weren’t answers

The first dashboard, committed on 8 November, had progression charts per exercise and weekly set counts. It was useful, and left all the interpreting to me. The direction I gave was blunt: “I don’t want to interpret all these numbers myself. I want my training questions answered.”

So questions came first. Fifteen became dashboard cards, starting with “Am I actually getting stronger?” In the first prototype, Claude wrote the answers, after I’d reviewed the evidence and agreed to send it. They kept failing the app’s own check that every number comes from the evidence: on synthetic data, 3 of 8 were rejected. The evidence lacked the comparison the question needed, and the model wasn’t allowed to state new numbers. I asked: “Why do we need an LLM to answer questions that could potentially be answered through reproducible calculations?” Python replaced the prototype and calculates the answers. Claude is now optional, never automatic.

Fifteen cards were also too many: only two could be answered, and several overlapped. I cut them to six and kept the rest as research.

Correct, but unreadable

The calculated answers were right and hard to read: ISO week codes, rep patterns, comparison windows and exceptions in one paragraph. Each answer now opens with a short insight in plain language, then two or three key figures, then rows that expand. Weeks are “last week” or “2 weeks ago”. The method and the week codes are still there, one click down.

Twenty-seven fewer sets

Then the workload answer for my own week said I’d recorded 29 sets, 27 fewer than the 56 the week before, and explained the drop exercise by exercise: fewer sets here, an exercise gone there.

I knew something it didn’t. I hadn’t trained on Friday.

I had the records checked session by session. The week before: Monday 12 sets, Wednesday 17, Friday 12, Saturday 15. The newer week: Monday and Wednesday with the same counts, and nothing after. The drop was exactly Friday’s 12 and Saturday’s 15.

Nothing was miscalculated. The app counted sets per week and per exercise, exactly as defined, and the tests checked those calculations, not whether they explained the week. GymBook’s records already carried each set’s day and routine. The analysis had never grouped those records into sessions. So it couldn’t tell fewer sessions from fewer sets within them.

The fix kept the weekly totals and added recorded strength sessions beside them: one routine on one day, matched between weeks by routine, not by weekday. The parts have to add up exactly to the weekly change.

What the records can’t say

The agents wrote the code. Changes went through backlog tasks, synthetic test databases and checks on my real data that printed only counts. By the third day the suite had 157 tests, among them tests for a missing Friday and Saturday session.

The app now says no Friday or Saturday session was recorded. It doesn’t say I skipped them. A missing record could mean I didn’t train, or didn’t log it. The records can’t tell those apart, so the app doesn’t try.


Share this post:

Previous Post
Running AI Locally Was Cool. Using It Was Another Story.
Next Post
That Thing You Keep Wishing Existed