How to Run a Screen-Time Audit on Yourself (the 5-Second Method)
7 min read · TimeLeak
Most people guess where their workday goes. The guess is always wrong, because the expensive parts of a day are invisible at the moment they happen: the fourth check of the same dashboard, the 15-second mouse journey a hotkey would erase, the "quick look" at a page you already looked at 40 minutes ago.
The fix is not discipline. It's data — and the data is almost free to collect.
The 5-second method
Run a watcher that records two things every 5 seconds:
- One line of metadata — active app, window title, idle time. A full workday is ~8,640 lines of text. Text is essentially free to store and nearly free for an AI model to read.
- One small screenshot — downscaled, local-only, auto-deleted after a week. You'll only ever sample ~20 of these, for the moments text can't explain.
What to compute (no AI needed yet)
| Metric | How | What it exposes |
|---|---|---|
| Minutes per app | count samples × 5s | where the day actually went |
| Context switches/hr | count app+title changes | thrash — most people land between 15 and 40/hr |
| Polling loops | same title visited 5+ times, short dwell | pages you re-check instead of batching |
| Longest focus block | longest unbroken same-context run | your real deep-work capacity today |
Then let a cheap model read the day
Feed the aggregates (not the raw frames) to an inexpensive model and ask one question: "what low-value behavior repeated today, and what's the concrete fix?" A day of metadata costs roughly a cent to analyze. The output you want is an observations ledger: each suspected leak, with a counter. When the same pattern shows up three days running, it's confirmed — fix it.
The rules that make it work
- Local capture only. Screenshots never leave your machine. This is non-negotiable — it's your screen.
- Confirm before changing anything. One weird Tuesday proves nothing. Three days proves a habit.
- Every finding must name a fix — a hotkey, a script, a digest, an extension. "Be more focused" is not a finding.
TimeLeak packages exactly this loop — the watcher is free, and Pro writes the daily brief for you.
FAQ
How much disk space does 5-second capture use?
Downscaled JPEGs run 100–200MB per workday. With 7-day retention that's under 2GB steady-state — the metadata itself is a few hundred kilobytes.
Won't watching myself change my behavior?
For the first day, slightly — then it disappears. The data you want is the repeat patterns, and those survive observation.
Do I need an AI model to benefit?
No. Minutes-per-app, switches-per-hour, and polling loops are pure counting. The model earns its cent by naming the fix, not finding the number.