# Skill: Build a personal AI OS -- persistent memory, specialist agents, and scheduled tasks from plain files

> Use this skill when recurring work (daily briefings, multi-project context, repeated role-switching) makes a single amnesiac chat session inefficient -- build a file-based "office" that survives restarts.

Source: sodigi-learn -- vibecoding/lichnaya-ii-os - https://sodigi.io/learn/vibecoding/lichnaya-ii-os

## When to use
- You re-explain your project context to the AI every session and are tired of it.
- You have repeated daily or weekly tasks (briefings, status checks, inbox triage) that should run automatically.
- You need multiple specialists (developer, copywriter, researcher) and want clear role separation.
- You want the AI to accumulate knowledge over time rather than starting blank each morning.

## Core rules
- Memory lives in files on disk (markdown + JSON), never only in the chat. The model re-reads files each session; it retains nothing in its "head."
- The whole system is four folders: core (CLAUDE.md), agents/ (staff personas), .claude/commands/ (slash commands), data/ (memory).
- Keep CLAUDE.md under 200 lines so it fits into context cleanly at every session start.
- One agent, one job -- a developer agent and a copywriter agent are separate files; one jack-of-all-trades agent gets confused.
- Scheduled tasks require an EXTERNAL scheduler (cron, systemd, LaunchAgent, pm2). The built-in session timer dies when the window closes.
- Logs are append-only -- never rewrite yesterday's log; only add today's entry.
- Secrets (API keys, passwords) never go into agent files -- use environment variables exclusively.
- At the end of every session, append a short log entry: what was done, what is left.

## Procedure
1. Create CLAUDE.md as the director persona: who it is, a table of specialist roles (name / scope / trigger verbs).
2. Create agents/ with one text file per specialist: their persona, which data/ files to read at startup, what they are not allowed to do.
3. Create .claude/commands/ with slash commands for recurring tasks (e.g., /daily-sync).
4. Create data/ with subfolders: daily-logs/, projects/, decisions/, inbox/.
5. Write a project context file in data/projects/ so the AI never needs re-orientation.
6. Wire /daily-sync: read latest log, check status, read inbox/, write morning summary to daily-logs/.
7. Schedule /daily-sync via the OS external scheduler -- not via a session-level timer.
8. At session end, have the agent append a log entry automatically.

## Ready-to-use prompt
```
Set up the seed of my personal AI OS in this project. Do it step by step:

1. Create CLAUDE.md -- describe: you are the director, you hand out tasks.
   Add a table of three roles: developer, copywriter, researcher -- and when to call each.

2. Create a data/ folder with subfolders daily-logs/, projects/, decisions/, inbox/.
   In data/projects/, put a short description of my project so I never have to explain it again.

3. Create a daily-sync command in .claude/commands/.
   It should: read the latest log, check project status, read inbox/, write a morning summary to daily-logs/.

4. Keep memory in files only. Do NOT write passwords or keys into any files -- use environment variables.
   At the end of the session, append a short log: what was done and what is left.

Show me the folder structure when you are done.
```

## Pitfalls
- Keeping memory only in the chat -- after a restart it is gone; everything important belongs in data/.
- Making one mega-agent for everything -- it gets confused; split into specialist files.
- Bloating CLAUDE.md past 200 lines -- the context fills before any real work starts.
- Using the built-in session timer for scheduling -- it dies when the window closes; use an external scheduler.
- Writing API keys or passwords directly into agent files -- one leaked file exposes everything.
- Skipping the log-append step -- without it the system never grows and forgets everything again.
- Dragging in a database for a handful of notes -- markdown and JSON are sufficient for a single user.