I built my personal system eighteen times
Between September 2025 and August 2026 I started eighteen personal productivity systems. Obsidian vaults, a Roam clone, React apps, single-file HTML apps, a SQLite database with a command-line tool. I'm a software engineer, so most were real builds with a schema, commands and git history, sometimes a small manual.
In August I finally did what I should have done at attempt three. I pointed a set of AI agents at the folders themselves: 21 repositories, about 1,100 commits, 339 coding-session transcripts and my task database. I asked them what actually happened. This is what they found.
The numbers
- The median system lasted four days. Half died inside 72 hours.
- None reached day 28.
- I started a new one every 3.8 days.
The system was the work
The details were worse than the averages. One app's own time tracker showed that a third of all the time it ever recorded was spent building the app. In another, the task I'd put the most time blocks against was "update this app". One vault had fifteen task-management plugins installed, tasks named "new one" and "new one-2", and a task timer that recorded zero tasks. A timeline script I'd written, twenty kilobytes of code, had logged two lines. Next to one system sat 704 KB of notes from books about productivity and 14 rows of actual use. Reading about the system beat running it by about fifty to one.
The folders also showed me repeating myself without noticing. The final version in one line of systems went back to plain dated log files, the exact layout I'd specified nine months earlier and then built away from. Another system's founding document names "building instead of running" as the first way it could fail. The same day, it made 135 commits.
What survived
Two things, and neither has a schema.
Things 3. I've used it continuously since September 2025. Every system I built to mirror it or replace it died. The app itself never did.
An append-only log. A date, then whatever I felt like writing. No checkboxes, no fields, no cadence. That same habit outlived the system around it five separate times, in different tools and years. One log kept going four months after every structured part of its folder had stopped. Another survived being moved between systems, so the practice travelled and the scaffolding didn't. In one folder, without any template or command, I'd written "I love just typing stuff to myself."
Anything with completion state (a habit tick, a routine checkbox, a weekly review) died. Thirty routine boxes, none checked. Habits ticked nine times, ever. Anything I could just append to survived.
The prediction I'd already made
One of those systems had a predictions page. It held exactly two predictions. One read: "p = 40%. Am I still using this in one week's time for planning? I have historically kept moving systems, but I want this to stick." I marked it skipped. I'd moved on to another system before the week was up.
What I'm doing instead
Not building a nineteenth. The third generation of one line has a rule written into its constitution forbidding a fourth. The folder I now use for coaching sessions with Claude opens with "Rule zero: do not build anything here." When it feels like the system isn't working, the answer is to take something away, not add a feature.
The one structured thing that ever compounded was somewhere else: a ledger of 25 pre-registered market experiments. Most of them were killed, one because the prediction had the wrong sign and another because of a bug in my own code. Writing down what I expect before I look is the only mechanism in eighteen attempts that produced knowledge I kept. I've never aimed it at my own life.