There is a difference between asking an AI to write a post and building a system that writes in your tone, assembles the artwork in your brand and still delivers the material to the DMs of whoever comments. The first part became a commodity. The second is where the work lives. That is what I put together over the last few nights, inside my own site, and I want to open the bonnet.
Why a machine, and not a prompt
Marketing run in a panic is what I criticise most. It made no sense for me to produce content in a panic. The idea was to invert it: one topic a week, three formats coming out of it (a LinkedIn post, a carousel and a reel), all inside a 52-week editorial calendar. The system knows the current week, knows the reference that anchors the theme, and knows the keyword that triggers the Instagram automation. What used to be a notebook of ideas became a production line.
The central piece is simple to describe and tedious to build: from the topic to text ready to publish, without me rewriting from scratch every time.
Getting AI to write in my tone
Any model writes correct text. The problem is that correct AI text has an accent. Too many em dashes, trios of adjectives, the gerund of false depth, the motivational sentence at the end. If I published that, everyone would smell it.
So the generation does not end at the model. It goes through a filter. First, a profile that describes my voice with explicit rules about what is forbidden. Then a deterministic post-processing step that kills whatever survived, with special attention to the em dash, which is fingerprint number one of generated text. The model suggests, the filter corrects, and only then does the text reach me. It took three rounds to get right. In the first two, clichés still slipped through.
Images in my brand: a reference is not pasting the photo
The visual part taught me the most. I wanted the cover of each carousel to be a new image, generated on the spot, but inside my identity. I tried three ways.
In the first, the system pasted the photo I chose into the background. Wrong. That is wallpaper, not generation. In the second, the AI created from a text description of the style. Better, but generic. In the third, which is the one that stayed, I send the reference photo to the AI and ask for a new scene in the same style. It looks at the light, the colour, the mood and the composition and creates something else that is similar in spirit, not a copy.
The detail that almost sank me: I had written into the instruction, without noticing, an order to keep everything in black and white. When I sent a colour reference, the AI discarded the colour and returned B&W. It was my own bug, one line long. I removed the lock and the reference started to genuinely lead, in colour and in light. On top of each image goes the brand, rendered on the server itself: the lime band, the title, the signature. No browser, straight from the function.
The part that turns into followers
Content without capture is entertainment. Since I have nothing to sell for now, the goal is straightforward: grow the audience and position myself. So the comment automation does not push product, it delivers value. Whoever comments the word of the week receives two DMs: a request for interaction, which the algorithm rewards, and a real diagnosis, in my method, with a light invitation to follow. The AI helps me write both, in my tone, from the content of the post.
The night the site "went down" (and had not)
In the middle of the build, the site stopped opening for me. It looked like a disaster. It was not. The server was responding normally to everyone outside. The problem was in the path between my machine and the hosting, in a new IP range that was handing dead addresses to anyone asking from Brazil. The fix was in the DNS, pointing the domain to the stable range. I note the lesson because it holds outside technology too: before assuming something broke, check whether the problem is the system or just your window into it.
What is coming
What is still missing is measuring results automatically, connecting to paid media when it makes sense, and the hardest part of all, which is not code: consistency. The machine is ready. Now it is a matter of feeding it every week. This text, by the way, was its first test in behind-the-scenes mode.