I just wanted it to start up lighter
Every morning, when my AI opens a working session, it reads a small note of instructions: "before you start, go review these things." It's how I give it context from the very first second. The trouble was that this note made it reread three whole manuals from top to bottom, every single time, even when the day's work had nothing to do with two of those three.
It's like forcing a colleague to reread the entire company handbook every morning before answering one email. It works, but it's wasteful. I wanted it to start light: read only an index of what exists first, then open in full only the page it actually needs. The way anyone uses a book: the table of contents first, then the right chapter.
What was already there, without my remembering
I asked the AI to find me a solution. The first thing it did was dig through the archive. And there came the surprise: the mechanism I wanted already existed. In my archive there were two system notes, written some time before, designed for exactly this: one is the index — given a topic, it says which drawer to look in; the other is the reminder of things to always know, holding the few rules to review only before certain specific actions.
So the problem wasn't a missing system. The problem was that the startup note ignored the two index notes and stubbornly kept making it reread the three whole manuals. It was rowing against the archive instead of using it.
Instead of guessing, it measured
This is usually where the thing I don't want happens: the AI proposes three different versions on a hunch, and you pick whichever sounds best. I stopped it right away. You measure, you don't guess.
We ran a real test, with a question that actually comes up for me: where do I add a certain automated task on one of my servers? I asked it both ways — with the old note that makes it reread the three manuals, and the new way that reads only the two indexes — and I measured how much material the AI was carrying at startup before it even began to answer.
The result
- What it reads at startup — old way: three whole manuals; new way: two indexes.
- Material loaded for nothing — old way: three times as much; new way: one third.
- Quality of the answer — old way: identical; new way: identical.
The new way carries 69% less material at startup, and the final answer to my question is identical, word for word. Translated: it starts lighter, costs less, and loses not a gram of precision. The old note simply made the AI work three times as hard for nothing before it was even useful.
The part I didn't expect
There's a detail that made me smile. The answer to the test question — where to add that automated task on the server — wasn't invented by the AI, and I didn't have to supply it. It came out of the archive: it was written in a note from months earlier, complete with times, paths and script names. The AI found it in two searches.
So, in that very same test, Timo did two things at once: it provided the data to decide how to make the AI start up lighter, and it provided the concrete answer to the work question. The tool to improve Timo was Timo. The memory I use every day improved the way that same memory gets read.
What I learned
Two things, simple ones. The first: you don't decide on a hunch, you measure — and almost always the measurement says something different from what "seemed right." The second, more important: most of what you need to decide well is already written somewhere. The value isn't having a brilliant AI that invents; it's having an archive where things stay, so that whoever reads them — me, today's AI, tomorrow's AI — starts from the right place instead of from scratch.
I took the light version. It took me a few minutes, because all the hard part was already there.
The same thing, ready for you
This startup note — the new, light one — isn't a secret of mine. It's the first thing I give my AI, and I prepared it so anyone using Timo can do the same, copying it as is.
→ You'll find it ready, with setup instructions, on the prompts page: /en/prompts/
In short
I just wanted my AI to start up lighter. Instead of guessing, I measured: 69% less weight at startup, the exact same answer. And the solution, like the test answer, was already written in the archive. Timo improved Timo.
— Rodolfo