What I Got Wrong (And Still Get Wrong)

The most useful version of this post is the honest one.

Here is what went wrong, roughly in the order it happened.

I Built Structure Before I Had Habits

The first time I tried something like this I downloaded a repo that had fourteen folders, a classification taxonomy, a set of templates with fifteen fields each, and a workflow document that ran to three pages.

I used it for two weeks before the structure became the obstacle.

The taxonomy required every note to be categorised before it could be created. The templates demanded completeness before anything was useful. The workflow document was the kind of thing written in an optimistic weekend and never consulted again.

The lesson Gall’s Law teaches is that complex systems evolve from simple ones that work. I tried to design complexity from scratch. Cerebro started with 1 note, 2 folders (Work and Personal) and no templates. That one survived.

The Capture Trap

Before Cerebro, I treated the vault as a place to save things: articles, links, fragments of conversations, notes from meetings. The theory was that everything useful would be there when I needed it.

The reality was that I had a very organised place where things went to be forgotten.

Saving something is not the same as understanding it. Bookmarks are not knowledge. A note that says interesting — follow up is not a note; it is a deferral. The vault filled with material I had no relationship with, and when I needed to find something useful, I was searching through noise. And burning through tokens.

The discipline I actually needed was not capture — it was synthesis: stopping to ask what I understood from something, writing that down, and discarding the rest.

The Maintenance Debt

There were periods over the 3 months when I stopped maintaining the vault consistently.

Each time, the effect was the same. The daily notes fell behind. The people profiles went stale. The memory files that give Claude context became inaccurate. When I opened a session and asked the AI to reason over recent decisions, it was working with information two or three weeks old.

A vault that is not current is not just less useful. It is actively misleading. The AI does not know its context is stale. It reasons confidently over outdated information.

The system only works if it is maintained. That is not a design failure — it is a feature, in the sense that any tool requiring discipline is a test of whether you have that discipline. Luckily, maintenance is automatic and you can ask Claude to do it for you as part of the skills you build to operate the vault.

The AI Hallucination Problem

Early in the build, I trusted the AI layer more than I should have.

There were sessions where Claude referenced something as if it were in the vault — a decision, a conversation, a piece of context, a person — that was not there. Or that was there in a different form, with different implications, and the model had smoothed over the distinction. Not frequently. Enough to matter.

I learned to verify. The model is excellent at synthesis and pattern recognition. It is unreliable as a source of factual claims about specific vault contents unless you can check the reference. Trust but verify is not a cliché here; it is the operating principle.

The Identity Confusion

At some point, I noticed I was spending energy maintaining the productivity system rather than being productive. The vault had become something I felt responsible for — a project to keep current, a system to protect — rather than a tool that served my work.

The correction is simple but requires regular honesty: is this serving the work, or has it become the work? The answer should always be the former. When it is not, cut the complexity until it is again.

What I Still Get Wrong

I still capture too much before synthesising. I still fall behind on meeting preparation in weeks when the schedule is dense. I still occasionally over-engineer a new section of the vault before testing whether the simpler version would do the job.

These are not failures of the system. They are failures of discipline — which is a different and more useful thing to know.

Post 5: whether you should build one, and the honest criteria for deciding.