How my AI system runs social media for 10 brand accounts, and why a person approves every post

My system runs social media by splitting the work into small jobs: an orchestrator hands each job to one of 13 specialised AI sub-agents, and a reviewer that starts fresh checks the result. A person approves every post before it is scheduled, because a wrong post on a public account is hard to take back. In its first five weeks it shipped 160+ posts for 10 brand accounts this way.

Why 10 accounts are hard to run by hand

I built it for a Philadelphia hospitality entrepreneur who runs 10 brand accounts. Each account has its own voice and its own mix of platforms, such as Instagram, LinkedIn, TikTok and Pinterest. A caption that works on one account rarely works on another, and a long LinkedIn caption will not fit in a post on X.

Done by hand, it is the same cycle every week. Someone researches ideas, writes a version for each platform, designs a graphic, checks the facts and schedules everything, for every account. No single step is hard. The trouble is that it never ends, and in a busy week the checking is the first step to get skipped.

The usual shortcut is one AI chat window and one prompt. It is fast, but the posts tend to come out in one voice, whichever account they are for. That sameness is a large part of what makes a post read like AI. So the task was to keep 10 voices apart while the volume went up.

One orchestrator and 13 sub-agents, each with one job

The orchestrator works like a project manager who never does the work itself. It reads a request, picks the playbook and passes each job to the right sub-agent.

The sub-agents are separate Claude-based AI workers. Each has its own written instructions and tools, and starts a job knowing only what that job needs. Between them they cover these jobs:

• The researcher finds ideas and checks each one against what the account has already posted, so it does not repeat itself.
• The caption writer produces a separate version for each platform, in that account’s voice.
• The designer makes the graphic, using a real photo when the post is about a real person, place or event.
• The reviewer checks the finished post, starting fresh.
• The scheduler prepares each approved post for every platform.
• The analytics step reads the results and passes what it learns back to the writer and the designer.

Some of the 13 do smaller jobs. One only finds source photos and does not write or design anything.

I split the work this way because one set of instructions covering research, writing, design and scheduling gets long, and long instructions are hard to follow and hard to fix. A sub-agent with one job has short instructions, and when it gets something wrong I know which one to open.

That is also how the 10 voices stay apart. Before any writing or design job, the agent reads that account’s guides: who the audience is, how the owner talks, which words to avoid, and what the account has been corrected on before.

How a post is checked before a person sees it

Ordinary scripts run first. They check the mechanical rules: the length limit for each platform, a caption for every platform the account uses, and no stray notes in text that will be published. They contain no AI, so they give the same answer every time.

Then the reviewer looks at the post. It is a separate sub-agent that starts fresh, so the agent that wrote a post never grades its own work. It gets the brief, the finished post and the account’s rules, and it never sees the writer’s reasoning. Anyone who checks their own work tends to read what they meant to write instead of what is on the page, which is why a writer gets an editor.

The reviewer answers ship, fix or ask. A fix comes with a list of the specific problems, and ask means the call belongs to a person. It blocks a post only for things that could hurt the account: a wrong date, a misspelled name, a headline cut off in the graphic, the wrong logo. Matters of taste become notes for me. If a problem survives a repair attempt, the post goes to a person instead of going round again.

A person approves every post

After the reviewer, I go through posts in a dashboard, where I can edit a caption in place or pin a note on the exact spot of a graphic that needs fixing. Then the post goes to the account owner as an email with the graphic, the caption and one line at the end: reply to approve. Nothing is scheduled before that yes.

Why not let it post on its own? I count an automation that cannot be stopped as a liability, so I do not build them. A system that can post by itself is likely to post something wrong at some point, and the first to notice is often a follower. By then people have screenshotted it, and deleting it does not undo that.

The system also only knows what it has been told, so it cannot know that an event moved to Saturday this morning unless someone says so. And the name on the account belongs to the owner, who should be the one to say yes.

The rules also tell the system when to stop and ask. If it cannot tell which account a post should be scheduled to, it does not guess.

Take a story whose main claim has no named source, place or date, where every copy points to another repost. The system stops before writing anything and brings it to me with options: share it as an idea, switch to an angle that can be checked, or drop it. I wrote that rule after a post where the system carried on instead.

It does cost something. Somebody has to read each post, and that takes time. If you want a system that needs nobody, this is not it. I think the time is small next to one wrong post on a public account.

How corrections become written rules

The system makes mistakes, so the aim is that each one happens once. When a post comes back with changes, I do more than fix that post. After it is approved, the system looks at what changed: my notes, the edits and the earlier versions. It writes each correction down as a short, general rule in that account’s lessons file, with the date and a one-line example. The writer and the designer read that file before every new job.

For example, after a review of one account’s whole feed, I added a rule that no layout repeats back to back, so the feed does not read as one template over and over. A script enforces it before the designer starts.

A separate maintenance agent looks after the instructions themselves. When something goes wrong, its only job is to find the instruction file responsible and propose a small fix to that file, so the fix lands once, in the right place.

Results come back too. The analytics step reads how each post did and which caption styles and layouts worked, and the findings go into the rules the writer and the designer read on their next run. It only reads from the platforms. It cannot post, schedule or edit anything.

What it has shipped so far

In its first five weeks the system shipped 160+ posts end to end, from the first idea to a scheduled post, across 10 brand accounts. Each one passed the scripts and the reviewer and then waited for a person’s yes. The case study shows examples of the graphics it made, each following its own account’s rules.

What this means for your business

Can AI run several accounts without sounding like AI? It can come close, as long as the voice is written down and a person keeps correcting the result. So start with one account. The system runs 10 today, but I would start yours with one, because the first account is where you find out what your rules need to say.

The setup is the first step of the service: your voice, colours, rules and the mistakes to avoid, written down once so every post follows them. Without that, the posts tend to sound like everyone else’s.

You do not need to change your tools. Posts reach you for approval by email or in a simple dashboard, and go out through a standard scheduling tool. Expect to give the most notes at the start, because each note is written down as a rule and does not have to be given again.

If you post on several accounts and want to see what this would look like for yours, the AI social media automation page lists what the service includes.

If you would like to talk it through, book a 30-minute call.