Telemetry — Rhythm
What my laptop knows about my week
I have a small program on my laptop that writes down what the machine is doing every 322 seconds and sends it somewhere I can keep it. It has been running for 138 days, which works out to 29,902 readings.
No single reading is interesting. It was awake at 3pm on a Tuesday in April. Four months of them stacked together show things that no single day does: when the work actually happens, how much the weekend really differs, and roughly when I go to bed.
Readings through rolled up page built
The week has an unmistakable shape
Each square below is one hour of one weekday, coloured by how often the machine was actively being used during that hour across the whole history. Darker means more.
The band across the small hours is the most obvious feature: from about 02:00 to 07:00 it is essentially never in use, seven days a week. The quietest hour of all is 05:00, active just 1% of the time; the busiest is 10:00 at 78%. Weekends are visibly paler than weekdays, and Monday and Tuesday evenings run late in a way the rest of the week doesn't.
A chart like this needs a lot of data to mean anything. Every one of these 168 squares has at least 101 readings behind it. Build the same picture from a day and a half of history and each square holds one or two samples, and it looks exactly as convincing.
Weekends are real, but only just
One bar per day, showing hours of active use. Weekends are greyed out. The typical weekday runs 11.5 hours against 6.1 hours on a Saturday or Sunday — so there is a weekend, but it is a shorter working day rather than a day off. The longest single day in the record is 19.1 hours, on 2026-04-16.
The work comes in long blocks, not fragments
Group the active readings into uninterrupted runs — stretches where the machine stayed in use with no idle gap — and you get something like a log of focus sessions. There are 1,119 of them across 138 days, about 8.1 a day, adding up to 1,348 hours.
The typical session runs 43 minutes, but the average is dragged up to 72 minutes by a long tail — the single longest unbroken run was 11.4 hours.
The chart below splits active time rather than session count, by how long the session lasted. Short sessions are common but barely register, because they hold so little time between them. 80% of all focused time lands in blocks of an hour or more. The day is a handful of long stretches with short ones scattered around them.
Each bar is the share of total active time spent in sessions of that length.
Week by week, it's been tapering off
Total focused hours per week. It ran near 94 hours a week through the spring and has come down since — the first half of the record averaged 75 hours a week against 54 in the second, and the typical session got shorter alongside it.
And it's a midweek thing
Average focused hours by day of the week. Wed is the peak at 12.4 hours, and the blocks are longest early in the week — Mon sessions run a median of 61 minutes. Both the amount and the length taper toward the weekend, which bottoms out around 5.7 hours.
A locked screen is a decent sleep tracker
It has no idea when I sleep. What it does record is whether the screen is locked. Take the longest unbroken locked stretch that overlaps the middle of the night, and you get something that behaves a lot like a sleep log — 85 nights out of 138 days, the gaps being nights I left the machine unlocked or away from it entirely.
The median night is 7.6 hours, running from about 1:09 AM to 8:36 AM. The wake time barely moves all week. Nearly all the variation is in when I stop.
Which produces the least comfortable number here: Mon nights are the worst, at 6.9 hours, and Sun nights the best, at 8.5. That's a 1.6-hour spread driven almost entirely by bedtime drift across the week, against a wake-up time that is essentially fixed.
| Night beginning | Asleep | Awake | Median | Nights |
|---|---|---|---|---|
| Mon | 1:16 AM | 8:25 AM | 6.91h | 5 |
| Tue | 1:42 AM | 8:52 AM | 7.03h | 16 |
| Wed | 1:02 AM | 8:50 AM | 7.53h | 21 |
| Thu | 12:52 AM | 8:16 AM | 7.47h | 12 |
| Fri | 12:59 AM | 8:33 AM | 7.87h | 16 |
| Sat | 1:02 AM | 9:03 AM | 8.14h | 6 |
| Sun | 12:23 AM | 8:25 AM | 8.51h | 9 |
Some weekdays have only a handful of nights behind them, so treat the day-by-day numbers as indicative rather than precise. The overall median, drawn from all 85, is the solid figure.
Nights grouped by length, in hours. The distribution is tighter than I expected. The short nights aren't rare events; they're the routine.
Two windows account for most of it
1,477 hours of foreground use, grouped into categories. A terminal and a browser take 86% of it between them. The editor barely registers, which looked wrong until I remembered that the agents doing the editing live in the terminal.
Only the category is recorded here. A list of individual application names with timestamps is a meeting schedule and a social graph, so the names are mapped to buckets during processing and thrown away. Time spent at a locked screen is dropped rather than counted as anything.