A one-person service business that runs on a small VPS: Postgres, n8n, git, Telegram and Claude Code. Eleven prompts build intake, booking, quotes, an approval queue you answer from your phone, and, new in 1.1.0, a classifier that decides before Claude writes plus the weekly marketing loop.
What you build
| Part | What it is | What it does |
|---|---|---|
| Eyes | A Telegram bot | Morning brief, approvals, warnings, kill switch |
| Brain | Claude Code, run on a schedule | Reads the rules and the database, writes the brief, drafts what needs a human |
| Memory | Postgres plus a folder of markdown (company-os/) | Postgres remembers what happened; the folder says how the business works |
| Nervous system | n8n | Every inbound thing goes through one event-router; nothing acts on its own |
| Hands | n8n workflows, one per job | Reply, book, quote, invoice, chase, post |
| Decides (new in 1.1.0) | Laya, a small classifier on the VPS | Tags any text in about a second for no tokens; code counts and routes; Claude only writes |
Where this comes from
The map is a published diagram by Structure Webworks, "The One-Person Company AI Operating System": forty jobs in seven departments, one closed loop from marketing to finance and back. Their stack is SaaS (HubSpot, ServiceTitan, QuickBooks, Zapier). This pack keeps the map and puts it on a box you own: Postgres, n8n, git, Telegram and Claude Code on a small VPS.
Version 1.1.0 adds the marketing half, and it follows two newsletters by re-structure.ai: split the jobs, connect them through one hub every tool reads and writes, keep a human gate on anything public or paid; and, from their Jev article, a small model that only decides, code that counts, a language model that only writes, a human who approves. The deciding part here is Laya, an open classifier (Apache-2.0) that runs on a CPU.
What you build
When it is running, a message from a prospect becomes a lead row, a drafted reply and a Telegram ping in under a minute. Quotes price themselves when the job is standard and ask you when it is not. On Monday, one file names the hook your competitors are not using, Claude drafts five posts from your rules files, a gate reads them, and you approve from your phone. Spam never reaches Claude.
The one thing it deliberately does not do: publish, spend, or send anything a human did not approve. Every send node sits after the approval check, and a kill switch in Postgres stops the whole thing with one row.
Do you need this? Be honest first
For a one-person service business that already has customers and loses hours to intake, booking, quoting and chasing. For someone who wants to own the system, read the rules it follows, and change them in a text file.
Not for a business that does not exist yet, and not for anyone who wants it built for them. If a notebook and a calendar are working, keep them. Prompts 7 to 10 only pay off once prompts 3 to 5 are answering real messages.
What you need
- A VPS: 2 vCPU and 4 GB for prompts 0 to 6; 8 GB if you want the classifier up all the time. Any provider with Debian or Ubuntu.
- Claude Code, logged in, on your laptop and later on the VPS.
- A Telegram account. The bot takes two minutes.
- For the receptionist, one channel you bring yourself: WhatsApp through a self-hosted Evolution API, or SMS through Twilio.
- About three hours for Phase 0, then one evening per prompt.
What is in the zip
one-person-company-os/
README.md setup from zero, the eleven prompts, what was run and what was not
COSTS.md every cost, with the date it was read
TROUBLESHOOTING.md the errors hit so far, with fixes
CREDITS.md where it comes from, what was verified, terms
prompts/00-06 interview, stack check, morning brief, receptionist, booking, owner queue, quotes
prompts/07-10 decision desk, weekly research, posts through the gate, triage first (1.1.0)
starter/ docker-compose, schema, n8n event-router and Telegram alert, setup scripts
starter/company-os/ thirty rules files: the folder that is the business
starter/desk/ desk.py, four recipes, sample data (1.1.0)
starter/laya/ the classifier's Dockerfile (1.1.0)
assets/brain-map.png the forty jobs on this stack
one-person-company-os-guide.pdf the same documents as one PDF
What was verified and what was not
Verified on a real Debian 13 box with Docker 29 and Compose v5: the compose stack, the schema, setup.sh, the event-router on n8n 2.36.8, the kill switch, and the Telegram alert. For 1.1.0, on a Debian box of mine on CPU: Laya 0.4.14 as a service with one checkpoint and a 3 GB cap, desk.py's self-test, and the four recipes on the sample data. The speed and accuracy numbers in the prompts come from those runs.
Not verified: the prompts as a sequence (0 to 6 and 7 to 10), the morning brief under cron, and Laya inside Docker from the shipped Dockerfile. My box runs the same settings as a plain service. The first people through will find the gaps; their fixes go in TROUBLESHOOTING.md.
Read the prompt before you run it
Hand each prompt to Claude first and ask what it plans to build. Compare that to the BUILD section. A prompt costs a minute to fix. A wrong build costs an afternoon.
