I Turned Myself Into a Dashboard for 30 Days and the Data Was Humbling
Photo: developer desk with coffee cup and laptop charts on screen, via thumbs.dreamstime.com
I have spent a large chunk of my career building observability for other people's systems. Dashboards for queues. Traces for request paths. SLOs with error budgets. And somewhere in the back of my head, a thought I couldn't shake: I have vastly less visibility into my own productivity than I do into a microservice I've never touched.
So for one month, I fixed that. I treated my workday like a production service — instrumented, scraped, and graphed. This post is the writeup of the experiment: what I built, what broke, and what the data actually said (which was frequently not what I would have told you over coffee).
The Instrumentation Stack
I wanted this to be cheap to maintain, because a self-instrumentation project that takes more than a few minutes a day dies within a week. Here's what I ended up with:
- Commits and code churn: a small script that pulled from
git logdaily, tagged with repo and hour of day. - Context switches: a background process on my Mac that sampled the frontmost app every 30 seconds and logged it. Crude, but effective.
- Meetings and focus blocks: a plain-text log file where I jotted one line per calendar event. No fancy integrations.
- Coffee: yes, I logged the coffee. Timestamp and cup count, entered on my phone.
- Self-reported energy: a 1–5 rating, once mid-morning and once mid-afternoon. Twenty seconds of effort.
Everything landed in SQLite, and a tiny Python script pushed it into Prometheus via pushgateway, which fed a Grafana dashboard I titled, with great maturity, "Prod: Me."
The total build time was a weekend. The total daily maintenance was about two minutes. That ratio turned out to be the whole ballgame.
What the Data Said
Finding one: my context switch count was obscene. I estimated I switched between apps maybe 40–50 times a day. The sampler said 200-plus on a typical day, peaking above 300 when Slack was lively. I don't think I'm unusual — but seeing the number graphed next to my commit output was genuinely uncomfortable. The correlation wasn't subtle: on days with the most switches, my deep-work blocks were the shortest, and my self-reported energy cratered by 3 P.M.
Finding two: coffee timing was helping me and hurting me. I'm a two-cups-by-9 A.M. guy, always have been. The energy ratings told a different story: my mid-morning scores were fine, but the afternoon dip was worst on the days I front-loaded caffeine hardest. I experimented with shifting the second cup to 11:30, and the 2–4 P.M. energy ratings ticked up about half a point on average. Small sample, self-reported, take it with salt — but it's a change I'm keeping.
Finding three: meetings didn't cost what I thought they cost. I'd have told you meetings were my biggest productivity killer. The data said otherwise. The real killer was meeting adjacency — a 1 P.M. meeting wedged into the afternoon destroyed the surrounding focus hours, but a 10 A.M. meeting cost me almost nothing afterward. Total meeting hours barely correlated with commit output; meeting placement correlated strongly. I now guard my 1–4 P.M. window like it's a production deploy freeze.
Finding four: my best code happens before noon, full stop. Commits between 9 A.M. and noon were both more frequent and — by my own judgment of the diffs — more substantial than afternoon work, which skewed toward chores, docs, and small fixes. I'd half-known this, but the chart made it undeniable, and I restructured my week around it.
The Honest Caveats
Worth saying out loud: this is n=1, self-reported, and subject to every observer-effect in the book. Just knowing the sampler was running probably changed my behavior — the classic instrumentation problem, applied to meat instead of metal. Some of my "improvements" may be placebo. I don't care. The dashboard paid for itself the moment it made one invisible pattern visible.
If you want to try this, my advice is the same as for production observability: start with the one or two metrics you'd actually act on, automate the collection completely, and resist the urge to build the perfect schema first. Your first dashboard is a hypothesis, not a monument.
The repo with the scripts is going up on my projects page once I've scraped the shame out of the commit history. Turns out even that data tells a story.