I have a personal assistant that lives in my terminal. It manages my tasks, captures ideas, writes blog posts, searches my notes, and runs my morning briefing. Today I tore it apart and rebuilt it as something more interesting.
The old system was a single 227-line prompt that tried to do everything. Email, calendar, tasks, blog, project management, knowledge search. Every time I asked it to add a task, it loaded instructions for managing my Gmail. Wasteful.
So I split it into 8 focused skills, each one a specialist:
The trick is that shared context file. All 8 skills read vault-context.md as their first action. My vault structure, linking rules, tag vocabulary, formatting conventions. One file. Update it once, and every skill picks up the change automatically. No drift. No duplication.
Here is the actual terminal session where the architecture came together. Claude planned the split, I approved it, and it built all 8 skills in parallel:
But here is the part that got weird.
I gave the system a nightly cron job at 2 AM that analyzes how I used the skills that day, checks for inconsistencies across all 8 of them, and makes one small improvement. It commits the change to a private GitHub repo and pushes. When I wake up, my morning briefing tells me what it changed overnight.
It is a system that watches itself and gets better without me asking. The improvements are small. A better example in a skill prompt. A missing tag added to the vocabulary. A description tweaked for more accurate auto-triggering. Nothing breaking. Just slow, steady refinement.
I have been running AI automations since March. But this is the first one that modifies itself. My terminal now has opinions about how it should work, and it acts on them while I sleep.
Tomorrow morning I will find out what it thought needed fixing.