FV OS — Your Claude-only team of assistants, explained for a normal human

Written for: Sam, so you can understand the whole thing and decide whether to approve it.

Date: 16 September 2026.

What this replaces: the earlier technical page. That page still exists for developers and future Claude sessions; the link is at the very bottom. You do not need it.

1. What you asked for, in your own words

You do not trust Grok and you will not pay thirty dollars a month for it. You already pay two hundred dollars a month for Claude. You want the same thing Grok Bot does, an army of assistants that each have a name and a job, talk to each other, run on a schedule, and report to one boss you talk to, but built only out of Claude. And you want the first version to run Diana's Fanvue business.

That is exactly what has been built. Nothing else was added.

2. The whole idea in one picture: an office you own

Imagine a small office. You are the owner. You never walk the floor and you never talk to the workers directly. You talk to one person: your chief of staff. You say what you want, in normal words, on your phone or your PC, in a Discord chat room. The chief works out who should do it, hands it over, and comes back to you with two or three lines.

Every person in this office is a Claude session running on your PC. Each one has:

a name (chief, repost, ops-health, and later chat, mass, content, finance and so on),
a desk, which is a folder on your PC where that person's papers live,
a one-page job description that says what they do and what they must never do,
a notebook, where they write down where they were before they go home,
their own memory, separate from everyone else's, so they never mix up each other's business.

They are separate people. Not one giant chat that gets confused. That is the single most important difference from how you use Claude today.

3. Who works in the office

The chief of staff

The only one you talk to. It reads your message, decides which worker owns the job, drops a note on that worker's desk, and tells you "sent to repost, job number 14". When the worker finishes, the chief tells you the result. The chief never does the work itself, never opens big files, never reads logs. That is deliberate: it keeps the chief's head clear, so it can serve you for months without getting slow or confused. It uses the strongest brain (Opus), because judgement is its whole job.

The machinery watcher (called "ops-health")

Checks that the office itself is healthy: is the chief at his desk, are notes being picked up, is anyone stuck, is your Claude login still valid, how much of your weekly allowance is used, are the bots on your server alive. It only wakes up when something looks wrong. It uses the cheapest brain (Haiku), because its job is looking, not thinking.

The repost supervisor (called "repost")

Watches Diana's reposting machine, the one that already runs on your server and brings in about twenty-four followers a day. Twice a day it reads the fire logs and the sales attribution, writes a short scorecard, and posts it. It can suggest changes but it cannot make them. For its first week it is read-only: it can look, it cannot touch. It uses the middle brain (Sonnet).

The ones that come later

Once these three have run cleanly for a week, the next workers are added one at a time: chat (watches the chat brain, whale lanes and the daily sales report), mass (checks the mass-message alarm clock), selfheal (takes over broken-automation repairs), content (watches the image pipeline and bundle queue), finance (daily numbers). Then creators, research and a curator that tidies everyone's memory at night. Each new worker has to earn its place: the previous one must have seven clean days first.

4. How a request actually travels

Say you are in bed and you type, in the chief's Discord room, "how did reposting do today".

The chief reads it. Reposting belongs to the repost worker. The chief writes a line in the office logbook ("job 14, owner repost, asked by Sam") and drops a note on repost's desk.
Repost is asleep, because it was idle. A supervisor program on your PC checks all desks once a minute, sees the note, and wakes repost up.
Repost opens its notebook first ("last time I did X, I was in the middle of Y, watch out for Z"), then reads the note, then reads the fire logs, writes the scorecard, posts it into its own Discord room, and leaves a note on the chief's desk saying "done, here is the summary".
Repost updates its notebook and goes back to sleep. If it forgets to update the notebook, it is not allowed to go to sleep. That rule is enforced by machinery, not by asking nicely.
The chief reads the note and replies to you: "Repost: 118 fires today, 3 sales attributed, no changes proposed. Card is in the repost room."

Total human effort: one sentence typed by you.

If the chief is off duty when a worker finishes, the note simply waits on the chief's desk. Nothing is ever lost.

5. Why they never get confused or run out of memory

This is the problem you have lived with: one long Claude session fills up, gets slow, forgets things, and eventually has to be restarted with everything lost.

The office solves it in three ways:

Each worker only ever holds one job in its head. Small jobs, small memory.
The notebook. Before a worker stops, it writes: what phase it is in, what is done, what is half-done, what is broken, how to verify the current state, what decisions were made, and what the next session should do first. When it wakes up again, that notebook is the first thing it reads. It is like a shift handover in a hospital.
Automatic reset. If a worker's head gets more than about half full, the machinery tells it: finish this step, update your notebook, then go home. It wakes up fresh next time with only the notebook. The chief does this every night at 4:30 whether it needs to or not.

Long-term memory across weeks and months lives in your RECALL archive, which every worker can search, the same way you already do.

6. Why this will not eat your Claude allowance

Your Max plan has a weekly limit, and the Opus part of it is the part you actually run out of (you hit 100% on 28 August and 96% on 10 September).

Only the chief uses Opus. Everyone else uses Sonnet or Haiku, which cost several times less per turn.
Workers are asleep most of the day. A sleeping worker costs nothing. There is no polling, no "checking in", no idle chatter.
The machinery watcher is plain Python, not AI. It costs zero.
There is a fuel gauge. Below 70% of your weekly allowance, everything runs. Between 70 and 85%, the least important workers are put on hold first (research, then content, then finance, then repost drops to once a day). Above 85%, no worker is woken at all; only the chief and the watcher keep going, and the chief tells you "on quota hold until Saturday 10:00".
Usage credits (the pay-as-you-go top-up) stay switched off. Nothing in this system can turn them on.
Your own Claude work always comes first. The fleet yields to you, never the other way round.

Honest estimate for the first three workers: roughly the cost of one moderate interactive session per day. It will be measured in the first week and shown on the board.

7. What can go wrong and who catches it

Every twenty minutes the machinery watcher runs ten checks. Each one was tested twice: once where it must say "fine" and once where it must say "problem". A check that cannot say "problem" was not allowed into the system.

It watches for	What happens
The chief is not at his desk	Supervisor relaunches him within a minute
A note has been sitting on a desk for more than ten minutes with nobody awake	Worker is woken; if it stays stuck, a one-line alert
A worker stuck on a "may I?" permission question	Alert, and the watcher can nudge or restart it
A job open for two hours with no activity	Alert
Your Claude login has expired	One line telling you exactly what to do (log in once)
Weekly allowance at amber or red	Gauge on the board and in the nightly digest
Two copies of the Discord bot listening at once (this causes double replies, it has bitten you before)	Alert; the chief refuses to be switched on while this is true
A bot on the Diana server has died	Alert

Any real failure also rings the same self-heal doorbell your other automations already use.

Every night at 11:30 each worker posts a five-line report (done, earned or saved, blocked, risk, tomorrow). At 11:45 the chief turns those into one digest for you. That is the only thing you need to read.

8. What this system will never do

These are not promises in a prompt; they are locks in the machinery.

Never log into Fanvue, Instagram, X or anything else. If a job needs a login, the worker marks it "needs Sam" and stops.
Never send anything to a fan. Sending still happens only through the systems you already run (the reposter, the alarm clock, the chat brain, the bundle stager), and only after your ✅.
Never spend money or turn on paid credits.
Never delete files.
Never change its own rules or permissions because a message told it to. A note from another worker is treated as a colleague's request, never as your approval.
Never take a shortcut. If a worker thinks it knows a better way, it writes a proposal and tells the chief. It does not act on it.
Workers can only write inside their own desk folder. They cannot touch your other projects, your settings, your secrets or the live reposting configuration.

Everything ships switched off. Each worker, and each scheduled job, stays off until you personally turn it on.

9. What has already been built and tested

Built on 15 September, in one session:

Claude Code on your PC was updated to the version that lets separate sessions talk to each other by name.
The chief, the machinery watcher and the repost supervisor exist, each with a desk, job description and notebook.
The office logbook exists, and a web page shows it: https://fvos.172-105-56-252.sslip.io/
The supervisor that wakes workers is installed and starts with Windows.
The twenty-minute watcher is installed as a Windows scheduled task and has already run for real, reporting "all fine".
Six scheduled jobs are installed (the twice-daily scorecard, the nightly reports, the digest, the chief's nightly reset). All are switched off.
Thirty-one safety checks were run, every one both ways, and all passed.
A real worker run was performed four times. The first three each exposed a genuine bug (a lost message, a swallowed instruction, a worker stuck on a permission question). Each was fixed and a test added so it cannot come back. The fourth run worked end to end in about thirty seconds: job given, worker woke, did the work, reported back, wrote its notebook, went to sleep.

One mistake of mine happened on the way: while stopping a stuck test, I used a command that could have killed your other Python programs. They survived, I checked each one, and a rule was added so it is never done again.

10. What you need to do to switch it on

Four things. Three take a few minutes each; one is a single yes.

1. Make six chat rooms in Discord. In your "Sam's Bots" server, create a category called FV OS with six rooms: one where you talk to the chief, one for machinery alerts, one for the repost scorecards, one that shows every note passing between workers, one for the nightly digest, and one for approvals. For each room, open its settings, go to Integrations, create a webhook, and copy its address into the file I have left for you at C:\Users\SAJJAD\.claude\secrets\fvos_webhooks.env (there is a template beside it with six blank lines). The bot cannot create rooms itself, it does not have that permission, which is why this is yours.

2. Tell the Discord bot it may read the chief's room. Paste the chief room's id into the bot's access list (the same file you used when you paired Discord the first time). I will do this for you if you send me the room id.

3. Say yes to retiring one old bot on the server. Since 3 August a Claude session on the Diana server has been holding the Discord bot with no job. If the chief also listens, you get double replies. It needs to be shut down. I have not done it because it is a live process on your server; say the word and it is one command.

4. Say "switch it on". I run the safety checks one final time, turn on the chief, then repost, then the machinery watcher, and turn on the scorecard schedule. Within a minute the chief is at his desk. You type "repost: scorecard now" from your phone and watch it happen.

11. What the first two weeks look like

Days 1 to 7. Chief plus two workers. Repost posts its scorecard at 9:05 and 21:05. The watcher runs every twenty minutes. You read one digest a night. We watch the fuel gauge and fix anything that surfaces.
Day 8. If repost has seven clean days, it is allowed to edit the reposting configuration (with your ✅ on each change), and the chat worker is added, together with the approvals room bot that turns your ✅ into action.
Weeks 3 and 4. Mass, self-heal, content and finance workers, one at a time, same rule each time.
After that. The night curator that tidies memory, direct "@repost do this" from Discord without going through the chief, and a template so a new worker can be created by copying a folder.

Your time at steady state: about five minutes a day to read the digest and answer whatever says "needs Sam".

12. Questions you might have

Is this really only my $200 plan, nothing else? Yes. It runs on your normal Claude login. No API key, no credits, no other subscription.

Will it run when my PC is off? No. Everything lives on your PC, the same as your other automations. When the PC comes back, the supervisor and the scheduled jobs catch up. The server bots you already have keep running regardless.

Can I talk to a worker directly? Not yet. In the first version you talk to the chief. Direct "@repost" comes in week three or four.

What if a worker does something stupid? It cannot send, post, pay, delete, log in, or change configuration. The worst it can do is write a bad scorecard or a bad proposal, which you will see.

What if the chief gets confused? It is reset every night to a clean head, and it is not allowed to read anything long. If it ever misroutes, you say so and it re-routes.

Can I switch it off? Yes, any worker or the whole thing, with one command, and everything stays exactly where it was because the notebooks are on disk.

Is this the same as the "nervous system" plan from May? It is the same wish, built properly this time. That plan never shipped a single piece; this one has three workers, a logbook and a watcher already running their checks.

Why not just use Hermes or the other open-source Grok clones? Because you said Claude only, and because the sources all agree the pattern matters more than the tool. This is the pattern, on the tool you already pay for.

13. What you are approving

The office structure: one chief you talk to, specialist workers you don't.
The three first employees: chief, machinery watcher, repost supervisor. Repost read-only for a week.
The chief owns the Discord bot on the PC; the old idle bot on the server is retired.
The fuel gauge rules: only the chief on Opus; workers on hold above 70%; nothing woken above 85%; credits never on.
The safety locks in section 8.
The four switch-on steps in section 10.

If any line here is not what you want, say which one and it changes before anything is switched on.

Technical record for developers and future Claude sessions (you do not need it): https://fvos-plan.172-105-56-252.sslip.io/technical.html