For the past months I've been running a real multi-channel solo operation — blogs, short-form video channels, social accounts, digital products — with an AI staff built on Claude Code. Not "I use AI to write drafts." The staff runs the pipelines: it publishes daily, checks its own work, reads its own numbers, and files me a report.
This post is the org chart, the four design rules that made it actually work, and the twelve failures that taught me those rules.
The org chart
owner (me, ~30 min/day)
└── chief-of-staff — routes work, keeps the blocker list, maintains memory
├── cmo — supervises all channels, daily standup, reallocates priorities
│ ├── blog-lead — queue, publishing, SEO, diagnosis
│ ├── shorts-lead — scripts, TTS, captions, upload, retention review
│ ├── social-lead — per-account templates, official-API publishing only
│ ├── newsletter-lead — digests, list growth, deliverability
│ └── commerce-lead — listings, affiliate links, CTAs, checkout health
├── analytics-lead — the only agent allowed to state numbers
├── research-scout — trend radar, competitor teardowns
├── product-lead — packages proven knowledge into products, runs launches
├── qa-reviewer — last gate before anything ships
└── ops-monitor — verifies pipelines are ALIVE (live state, not logs)
Each of these is a small markdown file in .claude/agents/ — a name, a description of when to use it, and a system prompt with its duties and hard rules. Claude Code routes work to them automatically.
The four rules that made it work
1. One lead per automation line. My first attempt was one mega-prompt that did everything. It drifted within a week — blog-voice on social, social-voice on the blog. Twelve narrow agents with clear lanes beat one wide one, for the same reason job descriptions beat "just handle stuff."
2. Delegation is a written contract. "Handle my blog" is abdication, not delegation. Every line the staff runs autonomously has a one-page charter: the goal, the daily routine, what it may do without asking, absolute guardrails, and — critically — a standing list of things only a human can do (signups, payments, legal). The agent acts freely inside the fence and stops at the fence, because the fence is written down.
3. Facts only, enforced in the prompt. Every agent carries the same rule: a number without a live source is reported as "unverified." No estimated views, no "roughly," no invented revenue — ever. This one rule is the difference between a dashboard you trust and a dashboard that flatters you.
4. Check live state, not logs. My most expensive incident: a publishing session expired, and the scheduler kept logging "success" while publishing nothing for six days. Now every status check verifies the outside world: does today's post exist at its public URL? Does the auth work right now? The log saying success is explicitly not evidence. Trust artifacts, not exit codes.
Three of the twelve lessons (the expensive kind)
- Invented numbers are worse than no numbers. An early agent "estimated" a metric to be helpful; the estimate got treated as real and drove a bad decision. Hence rule 3.
- Platform ToS beats clever automation. An unofficial posting client triggered verification hell across accounts. Recovery cost more than the automation ever saved. Official APIs or manual — no third option.
- Blockers hidden in chat logs never get cleared. "I couldn't do X, a human needs to" said once in a long transcript = never seen again. Blockers are now a standing checklist surfaced in every report until cleared. My job shrank to clearing them — usually minutes a day.
What this feels like in practice
Morning: the ops-monitor has already verified every pipeline is alive (or filed exactly what broke and what one action I must take). The CMO's standup names at most three priorities, each justified by a number with a source. I clear blockers, make one or two calls, and the staff executes.
The owner's job compresses to four verbs: set goals, clear blockers, review reports, decide pivots.
I packaged the whole system — the 12 agent files, the delegation charter template, the brand base, the persistent memory setup, the scheduling recipes, and all 12 lessons — as a starter kit you can install on your own business in about 10 minutes: AI Employee Company — Starter Kit.
Launch pricing is a public ladder: it starts at $19 and rises $10 with every sale (the listing shows the live sales count, so you can verify it). Early is cheaper — that's the whole mechanic, no fake urgency.
If you just want the ideas, everything above is enough to build it yourself — the kit saves you the writing and the twelve mistakes.
Top comments (0)