Before, maintaining a note system would devolve into procrastination under the guise of configuring Obsidian. Now I'm already at the far next level, where I'm playing with system prompts for AI agents that help me work with it. Fortunately, these skills then transfer to work in a refined form, so I have an extra excuse for why I'm dealing with the system and processes, rather than useful things. I realize this myself, so don't take it as a recommendation, I like it and just think it's a cool idea.
In a previous post on this topic I wrote that Claude Code is my assistant for filling regular notes - daily, first of all, where I dump operational memory. To a lesser extent weekly and monthly with retrospectives of the week and month, respectively. Well, and writing separate notes by templates is easy. But actually there are many more tasks. The principles of maintaining this note system imply that it somehow helps me document life, answer some questions about it, and ultimately achieve goals in different areas. And so I iteratively came to the current concept, which already looks somewhat towards Black Mirror.
I thought that a team of two "people" isn't enough for me. What if I assemble a full-fledged team with diverse experts? Imagine:
- Cleaner, who offers to organize notes in the system according to my rules or does it themselves, because I'm lazy to sort ready notes from inbox
- Content Maker, who could suggest ideas for posts based on what's been happening to me lately and a bunch of unfinished drafts
- Mentor, who could highlight how my actions relate to long-term career goals
- Medical Consultant, who could, looking at health information, make reasoned recommendations for specific actions to improve it
- Psychotherapist, aka reflection specialist, whose essence is to look at all this and highlight some general patterns and make me delve into their essence ..and maybe others.
In understandable words: each of them is a custom agent in Claude Code, with their own prompt, referencing notes that give more context on a specific topic. And the default agent remains such an assistant-router who redirects the request to a specific specialist. That is, if I ask how much longer to take some vitamins or when to take a test, the question will go to the doctor. If I ask to come up with a plan for preparing for interviews, the mentor will do it, because they simply have more knowledge on the topic.
Not that I discovered America with this. Setting up some custom subagents for each role in the team like product, analyst, architect, coder - this is already basics from vibe-coders. But I enjoy thinking about this precisely beyond programming and work projects.
The next step, from which I had a real aha-moment - make them sync, like a real scrum team. In understandable words: a custom /sync command, which gathers a council of all these subagents so they can share their thoughts on recent events, and the main one collects this into a report in standard format. Well, and weekly/monthly retros now, naturally, happen in exactly the same way.
As a result, I get a lot of constructive stuff. They ask questions that help me supplement information in their direction. Almost without my help they document the reporting period so I once again remember and fix events in my head. They trigger a lot of thoughts in me. They remind me what I wanted. They suggest actions. And they roast me for what I didn't do but should have.
The downsides of this system are execution time, token consumption, and lack of proactivity. I've been thinking a lot lately about how to set up the process to run agents automatically.