I Spent a Year Building My Own System, and It Lost to a WeChat Message

Have you ever downloaded a budgeting app? Full of confidence on install, three days of diligent entries, and by day four you never opened it again.

I did a nerdier version of that: I spent over a year building my own note system, and it met the same end.

I work in tech. The judgments I form, the traps I fall into, the reasons I change my mind — none of that survives in a WeChat favorites folder. I wanted a knowledge base I could actually query, one an AI could search, so last year I started building it: frontend, backend, task queue, line by line.

Features kept piling up. I added search. I added entity linking. Only one thing failed to keep pace: how often I actually wrote anything into it.

At first I assumed the features were lacking. Search inaccurate? Add hybrid retrieval. Cannot find relations? Add an entity graph. That is the reflex of an engineer — hit a problem, suspect insufficient capability, then go build more capability.

After a few rounds the feature list was genuinely longer. But opening the commit history showed the last piece of real content had been written long before, on a curve that did not match the feature work at all.

A system's capability and a system being used are two unrelated things. A long feature list with an empty usage log is the signal worth reading, not which module is still missing.

What actually blocked me was the act of making one entry.

Opening my own system meant: launch the client, find the right entry point, decide whether this belongs under stray thoughts or decisions. Three or four steps, under a minute total. But a minute is enough to talk you out of a passing impulse, and once the impulse is ground away by those steps, it rarely comes back.

The cost is not visible in the moment; it accumulates. Later, reviewing a technical choice, I found no trace of why I decided that way, and had to reconstruct it from memory — which may well be wrong.

Recently WeChat launched a hosted knowledge base tool called WeKnora. You send a message in a WeChat conversation and it is stored, searchable. Two weeks in, my capture frequency is higher than my own system ever reached at its most active.

It is not smarter than what I wrote. It compressed "make one entry" into opening an app that is already open and sending one line. What decides whether a system gets used is usually not how capable it is, but how many keystrokes the act takes.

I did not delete my own system. I split it into two layers.

The capture layer goes to the lighter entry point — a thought arrives, it gets stored. My own system now only does the genuinely complex work: turning scattered material into a publishable draft, running the review, exporting. Those steps cannot be skipped anyway, and handing them to a generic tool would not make them lighter.

Before switching I did look at other paths. There are open-source content automation projects with feature lists longer than mine, and I reviewed two rounds of them. Their default path is "auto-generate viral posts, auto-pass detection, auto-publish," which is the exact opposite of my requirement that material cannot be invented and every sentence must be traceable. Using them would mean handing over judgment too, and reworking them costs more than rewriting. So I did not switch.

What I actually gave up was one step: making an entry. Creation, and the decision to publish, are still mine, handed to no tool.

Engineers fall into one trap easily: believing that because they could build a better tool, they should build it. But "better" was never a longer feature list. It is fewer steps to finish one thing. The judgment that took me over a year to buy is not that my engineering was poor, but that I assigned light work — note-taking — to a heavy system.

This split is only validated in my own single-user setup. For team collaboration or systems requiring an audit trail, lighter is not automatically better; that is a different ledger and I am not qualified to judge it.

The budgeting card you never kept up with was probably not a willpower problem. Opening the app, finding the category, filling in the number — those steps were too heavy. Judge a tool first by how many actions it removes, then by how long its feature list is.

Before building a whole system for something, ask yourself one question: do I lack a tool, or do I lack a lighter entry point?