Continuing the notes topic, I want to emphasize again that it's become for me the main entry point into neural networks. First of all because I always start conversations not from a clean slate now, but with basic context about me, which I also manage.
All sorts of web versions of ChatGPT and Claude are all aware of the problem of lack of knowledge about the user, so they try to make themselves a "memory" where they write random slapdash thoughts taken from conversations with them. But here this problem doesn't exist right away. And even if they're also generated by a neural network - with the /init command, this is already better than fragments from ChatGPT memory.
And the other day I also came to the obvious thought that this is exactly the place where it's ideal to organize your GPTs, Projects, or libraries of custom prompts, whatever you want to call them. Not in the context of some project, but again in the broad sense of any of your activity.
For example, I wanted to simplify article summarization on the internet - I created a custom /tldr command in which it's written in what form I want to get a summary of content by a passed link. I wanted to automate analysis of data from a smart bracelet - I created a /healthcheck command that knows what exactly to get from HealthKit database, what to look at, and how to provide me a report. Say I'm solving LeetCode problems - I created a /leetcode command, where I pass a link/text/problem number, and it doesn't throw code in my face, but teaches me to think, because that's what's written in its prompt. Or if I want to immediately record my voice recordings in regular notes - I'll also rather make a command, not a full-fledged script. Well, you get it.
In short, if you have some base of prompts or ideas for what to use a neural network for - all this can just lie as custom commands in this project. These aren't necessarily commands related to notes, it's anything. This was a non-obvious insight for me.
Even if there are no notes, still starting a project for an agent and putting all the goodies there is a great idea.
And if you're steamed about the portability of these commands between agents, write everything in agent-agnostic files, and reference them from commands. This works fine even without the commands feature if the system prompt has info like docs/Tldr.md - description of /tldr command. With web tools there's no talk of any portability at all.
Now I'm testing the idea of running the terminal with this agent in quake mode via hotkey. Seems like a rich idea - to raise this agent-secretary from anywhere, similar to something like Raycast. Liking it so far.