In short — the case at a glance
The owner of an expert Telegram channel stopped writing posts himself — and his followers never noticed, they read it just like before. He used to run it whenever inspiration struck: irregular, dependent on free time. I built a two-layer system that took content off his plate entirely. The first layer is planned: a neuro-agent profiles the audience, builds a weekly content plan, writes a post from it every day, runs it through an editor-agent, and publishes it. The second layer is reactive: a separate workflow monitors world media daily for the client's keywords and turns a fresh story into a post with a link to the source. The point of both layers is the same: the posts are written in the client's voice and are indistinguishable from the ones the owner used to write by hand.
This is my work as a fractional AI consultant — building neuro-agents that take routine off a person. Not 'magic AI that replaces everyone', but a concrete stack of services for a concrete job. I built the whole content cycle myself: the n8n orchestration, the prompts for generation and editing, and the plan storage. This is automation, not classic software development — and I say so plainly, not to inflate the price.
What it was like before
The client had a live expert channel but no system. The classic 'I'll write when inspiration strikes' — and inspiration doesn't run on a schedule:
- posts came out irregularly, driven by mood and spare time;
- there was no strategy or content plan — topics were invented on the spot;
- the process drained energy and returned little.
The ask was simple, but with a hard condition: the system had to generate and publish content on its own, yet the posts had to be indistinguishable from ones the owner wrote by hand. Automation that produced 'typical AI text' would have killed the whole point — people follow the channel for a specific person, not for a neural net's feed.
How the system works: two layers on one engine
Instead of manual writing, I built an architecture on n8n, Supabase, and GPT. It works like a small content factory with two shops — a planned production line and a rapid-response news desk.
Layer 1. The planned line — 5 steps
Audience profile
I broke down the target audience into segments: who the readers are, their pains, interests, triggers. This analysis is the fuel for the whole system — every post is written against it.
Automatic content plan
Once a week the neuro-agent builds a new publishing plan. A table appears in Supabase: topic, keywords, and the audience segment the post targets. A schedule a week ahead.
Daily post writing
At the set time n8n fires the workflow: it takes the day's topic from the plan in Supabase, generates a draft via GPT, and hands it off to the editor.
Editor-agent
A separate module works as an in-house editor: it cuts filler, sharpens arguments, and tunes the text to the owner's voice — based on his real posts, phrasing, and manner. This is the key step: here the post stops being 'text from a neural net' and becomes indistinguishable from what the author would write himself.
Auto-publishing
The approved text goes straight to the channel. No manual action. The owner can still adjust the plan or any published post at any time.
Layer 2. The news desk — 4 steps
Once the planned layer was running, we added a second one: live posts on the day's agenda, so the channel reacts in the moment.
News request
Every day a scheduled trigger fires: the system pulls a batch of fresh publications for the client's keywords via an API.
Material selection
A code node filters the news by topic, randomly picks one fresh item, and assigns it a random publishing time during the day — so the channel doesn't look robotic.
Post generation in the author's voice
The chosen story goes to GPT, which writes a short post in the first person: a comment, some irony, a tip for subscribers — in the same voice as the planned posts. Not a dry rehash, but the author's personal reaction, as if he'd read it and replied himself. It always ends with a link to the source.
Auto-publishing
The finished text goes to the channel automatically via the Telegram integration.
Both layers live on one engine and complement each other: the planned line holds regularity and depth, the news desk holds relevance.
What came out of it
Honestly, with no invented reach numbers:
- The posts are indistinguishable from hand-written ones. The main result: both the planned editor and the news workflow write in the owner's voice — the reader can't see the line between 'wrote it himself' and 'the agent wrote it'.
- The idea-to-publication cycle is fully automated — posts go out every day, steadily, week after week, without the owner.
- The channel feels alive: on breaking topics posts go out same-day, so subscribers feel the author is always current.
- The owner is freed up: the time that used to go into scanning feeds and writing posts now goes to product and sales.
- The system scales easily: layers are added one at a time — first planned, then news, then new sources.
Why this beats 'hiring a copywriter'
A copywriter writes while they're paid and while they have the energy. And you can almost always tell it wasn't you — someone else's phrasing, someone else's rhythm, 'sort of my ideas, but it doesn't sound like me'. A neuro-agent doesn't tire, doesn't miss deadlines, and doesn't lose the voice, because it's tuned to one specific voice from the start — the owner's. But that's not even the main thing. This isn't 'a bot that posts', it's a system with roles: one agent builds the plan, another writes, a third tunes it to the author's voice, a fourth watches the agenda. The owner stays the author and editor-in-chief but stops being the operator.
Why fully autonomous AI content falls short and where the line runs between machine and human — I covered that separately: pure AI loses, the hybrid wins →
Does your channel run on inspiration and spare time?
That's not a ceiling — it's the absence of a system. I build neuro-agents that run a channel on their own — in the author's voice, from plan to publishing. I'll look at your case and tell you honestly whether it can be built.
