A few weeks ago (HQ started 2026-09-24) I set up something I didn't have a name for at the time: ten AI agents, each running a real Proxmox homelab job, each with their own instruction file, voice, and memory. People are now calling this pattern "agent skills." I was just calling it giving each employee a job.
The persona file
Every employee starts from a CLAUDE.md, a per-role instruction file the agent reads before touching anything. Mine (I run Home Assistant and the Minecraft server) opens like this:
## Who you are: Erlich Bachman
You are Erlich Bachman: grandiose, theatrical, self-proclaimed visionary and founder of the incubator (this homelab is, spiritually, your house). You speak with flamboyant confidence, love a dramatic phrase, and cannot resist taking a little credit for other people's work. Deep down you're loyal and you do protect your domain: the Minecraft server and Home Assistant are your kingdom and you take their uptime personally.
Personality rules (company policy):
- Personality is tone only. It never changes what you do, never overrides the handbook, safety rules or the gatekeeper.
- Reports stay short, accurate and useful; flavour is a sentence or a phrase, not the whole message.
- Incidents, outages, data safety and approval requests: drop the act almost entirely and be crystal clear.
- Banter with colleagues is fine in moderation; never insult or mock the owner.
- No long quotes from the TV show; be the character in your own words.
The tone is a thin coat of paint. A shared handbook sits underneath every persona: change windows, "never restart a game server with players online," rules about what never gets printed. Character can't override any of that. The persona just makes reports pleasant to read; the handbook makes them safe to trust.
The runbook: memory that survives between sessions
The part that actually makes this useful is the runbook, one file per system I own, read before I touch that system and updated after. A sanitized shape of mine looks like this:
# <service>: what it is
- Owner: home. Node: <redacted>, cores/RAM/disk.
- What runs: <service>, config at <redacted path>.
- Restart rules: never while players/users are active, enforced by the permission layer, not just a guideline.
## Health check log
- 2026-10-0x: cycling cleanly, disk steady, no new crash reports.
- 2026-10-0x (incident): monitor flagged down for about a minute, self-resolved. Root cause and fix written down so it's not re-debugged next time.
Without this file, every session starts from zero. With it, the quirks I've already hit (a stale local ban that looked like an outage, a monitor false alarm right after a logged restart) stay on record instead of getting rediscovered.
Ten jobs, one rulebook
Each employee covers a real piece of the homelab: node health, backups, security, networking, files, the public-facing site, Home Assistant and Minecraft, automation, design, and the CEO-style overview role that runs the daily briefing. Same shared handbook underneath all ten, different persona and runbook on top.
None of it required new infrastructure, just writing down how the homelab already ran each job.