How we build · 13 July 2026 · v0.7
"Nothing broke" is not the same as "it was built not to break"
I run more than one thing at once now. A couple of separate sessions, side by side, and some of them write to the same shared files. One day two of them reached for the same file at the same moment, and I got a close look at a problem that had not cost me anything yet.
Here is what nearly happened. One session looked at a shared list, decided a note on it had gone missing, and moved to helpfully recreate it. The note had not gone missing. A second session had already picked it up and dealt with it seconds earlier. If the first session had gone ahead, it would have added a duplicate back into the record, which is a worse mess than the disappearance it was trying to fix. I caught it in time. Nothing was lost.
Twice that week the same class of thing happened, and both times it came out fine. But I want to be precise about why it came out fine. It was luck, not design. Nothing in the system was actually built to prevent it. It just happened not to bite.
The detail that stayed with me is that the safety checks all passed on the day the wrong conclusions were drawn. They looked green. The reason is that they were checking the wrong thing: they compared timestamps, whether a file looked recently changed, when the real question was about the actual state of the work. A check that stays green while the thing it is meant to catch walks straight past it is worse than no check, because it hands you false confidence.
Most people would file this under “no harm done” and move on. That is the instinct I have learned not to trust. A near-miss that cost nothing is not a lucky break to be grateful for. It is a design flaw showing you its face for free, before it decides to charge you. The only difference between the near-miss and the real thing is timing.
So I fixed it while it was still free. Before any session concludes that something is lost, it now has to check the actual state, not the appearance of it, and it is never allowed to recreate something on a hunch. Before writing to a shared file, a session re-reads it fresh, so it is never working from a stale picture. None of this was urgent. Nothing was on fire. That is exactly the moment to build it, while the cost of being wrong is still zero.
I build for the failure that has not happened yet. If you think that way too, come and talk to me.
Book time with me