I Tore Down Three Tool Setups in Two Months. What Was Left?

This morning, a little script I use for my writing threw an error: embeddings HTTP 404.

In plain words: the endpoint it wanted to call wasn't there anymore.

I hadn't touched the script. I wrote it in July against a hosted knowledge-base service. In August I stopped using that service and moved to a copy on my own server, and the door the script keeps knocking on doesn't exist in the new place.

It's a small aftershock from this summer's tinkering. The funny part: in two months I tore down three sets of tools, and what has followed me all the way to today is a handful of plain text files.

Some background on what I was tinkering with.

I write a WeChat blog on the side. The material comes from articles I read, pitfalls I hit, and trade-offs I made. That stuff was scattered everywhere and I could never find it when I actually sat down to write, so since June I've been building myself a pipeline for "collecting material and helping me draft", with a fair amount of AI in it.

In the first two months, the foundation of that pipeline changed three times.

The first was Trilium, an open-source note app that lets you pile notes into a personal knowledge base as a tree. I ran a copy on my own server. It went offline on July 7.

The second was an app I wrote myself. I started it on June 16: backend, frontend, background jobs, database, the whole set. It ended up as 102 code files and over twenty thousand lines. The last time I used it to save anything was also July 7. From first commit to last real use: 21 days.

At 3:04 pm on August 3, the archive for that app finished packing. Four containers and a database with 24 tables were deleted right after.

The third was WeKnora, an open-source knowledge-base framework from Tencent. You drop articles in, and it can search them by meaning, so you don't have to guess which keyword you used back then. I started on the hosted version. That same day its web interface broke, so I just set up my own copy on my server and moved all the data over.

None of those three teardowns happened because the tool was bad to use.

Looking back, why didn't those twenty thousand lines survive two months? It took me a while to see it: the tool, the environment, and the goal are three separate things, and you have to account for them separately.

Back then I treated "having a system of my own" as the goal. The real goal was one sentence: don't lose the good stuff I read, and be able to find it when I write. The app was just one tool for that sentence, and a heavy one.

Then the environment changed. When I started, I assumed the AI needed an interface and a dedicated database to feed it. Halfway through, I realized the AI coding assistant I use every day could already read files on my computer and call APIs directly. Nobody needed to pass through that web app layer anymore.

What collapsed was the premise the app stood on. I didn't see that coming at all; I can only say it in hindsight.

So what survived? A few Markdown files.

They say who I am, what this blog writes about and what it doesn't, which questions a topic has to pass first, and which sentence patterns are banned. The earliest persona file was created on June 9, written as reference material for a custom GPT inside ChatGPT. Later the writing work moved to a scheduled task on a different AI, and these files just kept working. I just checked the modified date: June 15. Untouched for three and a half months.

AI tools understand plain language, so the few pages that spell out "what I want" turned out to be the part of the whole setup least afraid of the next generation.

When I switched AIs, not a single word in those files changed. The new task simply reads them from where they already were.

Put the two side by side and the gap is hard to ignore.

The twenty-thousand-plus lines took more than a month of my spare time, and now they sit in a compressed archive on my server. The text files took about a week of edits, and they've been in use ever since.

Learning tools is obviously useful. That 404 this morning still needs someone who can read an API to track it down. The difference is shelf life. Effort spent on a particular tool expires when the tool is swapped out. Effort spent on writing down the goal carries over to whatever tool comes next.

AI tools now turn over every few months, so the order I've set for myself is: first write down, in black and white, what I want it to get done, then go pick a tool. If I pick the wrong tool, I just switch again, and what I lose is only the tool layer.

I've only tested this judgment for a little over two months. Whether it still holds two or three years from now, I can't say.

There's one signal I'm watching now: the modified date on those files. If the day comes when I find myself editing them to accommodate some new tool, that means I've tangled the goal and the tool together again.