For a long time I wanted to publish more. Then I noticed something uncomfortable: writing was only half of the job — and not the bigger half.

Pick a topic. Gather sources. Outline. Draft. Check every fact. Make a cover. Cut a video. Rewrite the same idea for Telegram, X, LinkedIn, Instagram and Medium — each with its own length, tone and format. By the time an idea was ready to leave the house, there was little time left for the idea itself.

So I stopped looking for a better chatbot and started building a system.

The work around the work

Most independent creators don't run out of ideas. They run out of hours. The bottleneck isn't inspiration — it's everything wrapped around it: research, formatting, checking, packaging, posting, measuring.

That's the part I wanted to hand off. Not the thinking. The friction that keeps me from thinking.

Not a pile of AI tools. One system.

I already had plenty of AI tools. The problem was that every session started from zero. Each chat forgot who I am, what the project stands for, which facts are sacred and which words we never use. The results were fast — and different every time.

So instead of adding more tools, I built one system where each AI agent has a single job, a clear brief and written rules. Picture a small editorial office where the staff are specialized AI operators, and every handoff follows the same route.

I call it Cirqus Mundi — the Order’s operating world. An axis at the center, the rings of the Order’s institutions around it, each with its own agents. That is the whole idea: the value isn’t in any single part. It’s in how the parts move together.

Why my first AI agent wasn't a writer

The obvious move would have been to start with a writer. I started with a coordinator instead — The Conductor.

The Conductor doesn't write posts or invent ideas. It runs the system: it plans and assigns tasks, schedules them, passes one operator's output to the next one's input, and routes every draft through the checks before it reaches me for the final call.

Only then did the specialists arrive — the Operators. Among them: a Scout that looks for topics, a Writer, an Artist for covers, an Editor for short videos, a Verifier that checks facts, a Publisher and an Analyst. I'll introduce them one by one in this series.

The rulebook matters more than the model

The biggest lesson so far: giving an AI a task is not enough. Without context and rules, you get a coin toss.

So the system has The Archive — a written canon and a set of regulations that every operator receives before it does anything. Who we are. How we speak. What we never promise. Which facts can't be bent.

A concrete example. On 31 October 2024 I was hacked and lost 1.088 BNB (the full story is here). The number is written "1.088", never "1,088", which an English reader would take as more than a thousand coins. One wrong comma would turn a painful lesson into an absurd claim. That rule now lives in the canon, and the Verifier checks every article against it before I ever see the draft.

What works — and what still breaks

I promised myself to show an experiment, not a polished success story. Here is the honest status.

What works

• Routine can be delegated. Formatting, adapting one idea for different platforms, scheduling and cleanup now run with almost no input from me.

• Rules make results stable. Since the canon was written down, drafts stopped drifting off-brand.

• A second pair of eyes catches real mistakes. A separate Verifier, working in its own context, is much harsher on a draft than the agent that wrote it. It caught several problems in this very article — including a date that didn't match the record.

What still breaks

• A draft is not a finished piece. The operators help with form. Meaning, accuracy and the final call are still mine.

• Platforms aren't built for this. One network froze my developer account for "unusual activity", and posting there went dark until I proved it was me. Some sites have no publishing API at all, so a human click stays in the flow.

• Agents make confident mistakes. My site-checking agent once reported that no article on the site had a headline. The articles were fine — the checker was broken. That's why a person reads every report before anything gets "fixed".

• AI images are unreliable with text. Letters get mangled, so cover text is now added by code, not drawn by the image model.

The one rule I won't break

Everything above follows a single principle written into the canon:

Agents extend the capabilities of the Order. They do not replace the judgment of its people.

I don't want AI to think for me. I want it to remove the work that stops me from thinking. Every article still waits for my approval before it goes live.

What's next

Cirqus Mundi is still being built. In the coming posts I’ll open it up piece by piece:

1. Cirqus Mundi: why I built my own AI system

2. The Conductor: who runs the machine

3. The Operators: the specialists inside

4. The Archive: why even good agents fail without a rulebook

5. One full cycle: from idea to published piece

6. Mistakes and limits: what AI still can't do well

Not an advertisement for neural networks — a working log of one person building their own infrastructure, and an honest account of what comes out of it.

Your turn: do you already trust AI with a whole part of your work — not just a conversation? What works for you, and what still doesn't?