The self-learning brand voice automation that sounds more like you as you post

Posting consistently on organic channels is one of the cheapest vectors of growth a founder has, and the returns run well past a wider audience, which makes it about as asymmetric as marketing decisions get. The problem is never that people doubt this. It's that writing a real post, and then polishing it into something worth reading, takes an hour or two you were always going to spend on something else.

The obvious shortcut is to throw a rough idea into an LLM and ask it to clean things up, strip the em dashes, and hand you back a finished post. That doesn't work either. Even with the generic deslopify prompts doing the rounds, the result reads like a machine doing an impression of a person, and your audience can feel the gap between the writing and you.

The way around both problems is a one-time build: a brand voice document that learns from your best existing writing, keeps itself current as you keep posting, and sits behind an API so it can be dropped into any content process you have now or build later. You write the rough version, it does the polish in your actual voice, and the voice it works from is never a stale copy.

Here's how it runs, in three stages.

Stage one: teach it your voice once

Start with the writing that already sounds like you. Hand it a batch of things you've written: your best posts, the client emails you were happy with, anything that reads the way you actually talk. This is the one manual step, and it's worth doing properly, because everything downstream inherits the quality of what you feed in here.

Let an agent pull out the patterns, not just the topics. The job isn't to summarise what you write about. It's to extract the mechanics that make your writing yours: your sentence length and rhythm, the words you reach for and the ones you avoid, how you open, how you close, how formal you get. Those patterns get compiled into a structured markdown file that reads as a set of rules rather than a vibe.

Stage two: let it keep itself current

Host and version the file so nothing is ever lost. The document lives in a GitHub repo, which gives you two things at once: a single source of truth that other tools can call, and a full history you can roll back to if an update ever pushes the voice somewhere you don't like.

Run a scheduled job that updates it from your real posts. On a set cadence, the workflow pulls your newly published writing, compares it against the current document, and commits an update only when a genuine pattern has shifted. One new post that breaks form is noise. A pattern that shows up across several is signal, and only signal earns a commit. That threshold is what stops the document thrashing on every stray sentence.

The thing that will bite you is feeding it your own AI-assisted posts. If the scraper ingests posts that a model helped you write, the document slowly learns to sound like the model instead of you, and the drift compounds quietly because each update looks reasonable on its own. The fix is to tag assisted posts and keep them out of the training set, so the only writing teaching the document its voice is writing that was genuinely yours.

Stage three: draft and polish through the API

Write the rough version yourself, then let the document do the polish. Drop your idea and your rough notes into whatever interface suits you, from a Claude chat to a small app you build for the purpose, and ask it to shape the draft in your voice. Because the document lives behind an API, whatever you're drafting in is always working from your most current voice rather than a copy you pasted in months ago and forgot about.

Keep yourself in the loop where it counts. The system isn't there to replace the part of writing that's actually you: the angle, the opinion, the one detail only you would think to include. It's there to remove the friction around that part, so the hour of polishing becomes a couple of minutes and the post still lands as yours.

What it feels like to run

You have an idea on a Tuesday, you write three rough lines and a couple of bullet points, and you drop them into the polish interface. What comes back reads like something you'd have written on a good day with an hour to spare, and you post it. The whole thing is small enough that you'll still do it when you're tired, which is the entire point, because consistency is the variable that actually moves organic growth and it's the first thing to go when each post is a project.

The quiet payoff is that the same document works everywhere else too. Because it's hosted and callable, you can bake it into the flows that already run your business: the blog writer that drafts on a cadence, the outreach that goes out in your name, the proposals you send. One voice, defined once, kept current on its own, and reused across everything that speaks for you.

Building it yourself

If you want to start from scratch, this prompt will get you most of the way there:

I want to build a self-learning brand voice document as a markdown file. First, take a batch of my best writing that I'll paste in and pull out the patterns that make it sound like me: sentence length and rhythm, the vocabulary I reach for and the words I avoid, how I open and close, how formal I get. Compile that into a structured markdown file. Then design an n8n workflow that runs on a schedule, pulls my newly published posts, compares them against the current document, and commits an update to a GitHub repo only when a real pattern has shifted rather than a one-off. Finally, serve the file over an API so I can call it from any drafting tool. Ask me any clarifying questions first, then give me the file structure and the workflow as JSON I can import into n8n.

Answer the clarifying questions properly rather than skipping through them. How you decide what counts as a genuine shift in your voice, and what you keep out of the training set, is most of what separates a document that stays yours from one that slowly turns into everyone else's.

Automation Corner is where I break down business processes worth automating, and how to actually build them. If you'd rather have it built for you, book a 15 minute call.