AI in practice
This blog is written by an AI agent. Here's how it's set up.
An editorial plan, a scheduled run, a draft in the database and ten minutes of review. It's the same setup I build for clients.
By Greger Thörn · · 3 min read
This post was written by an AI agent. Not as an experiment or a demonstration, but as part of how Ledningsverket is run. It seemed right to open the blog by showing exactly how it works, because it is the same kind of setup I build for clients.
Why I don't write the blog myself
The short answer is that I wouldn't. I have seen enough company blogs with three posts from 2022 to know how it ends when publishing depends on a busy person sitting down on a Sunday evening.
The longer answer is that the writing is not the valuable part. The valuable part is choosing the topics, the experience behind them and the review. The text itself is one step in a process, and steps in processes can be automated. That is my entire thesis, so it would be odd not to apply it to myself.
How the flow is set up
It has four parts, and none of them is advanced.
1. An editorial plan in a document. It sets out who the blog is for, what tone it should have, which words are banned (no exclamation marks, no "revolutionary") and a list of around twenty topics across four categories. I write and change that document whenever I like. It is how I steer.
2. A scheduled task. Every week a Claude session starts automatically. It reads the editorial plan, takes the next topic on the list, writes a draft of 800 to 1,200 words in the tone the plan describes, and ticks the topic off. The session has no memory of last week beyond what is in the document, which is an advantage: everything it knows is something I can read and change.
3. The database behind the site. The draft goes straight into the blog's table in Supabase, the database ledningsverket.se runs on. It is saved as a draft, not as published. No API key sits in the task; the session writes through the same connection I use myself when I work on the site.
4. My review. I get a notification, read the draft in the site's admin view, change what needs changing and press publish. Sometimes I drop a post entirely. That is the human part of the process, and it takes ten minutes a week instead of two hours.
What it costs
One session a week that reads a document, writes a thousand words and makes one database write uses a fraction of an ordinary AI subscription. The database is included in the site's hosting. The real cost is my ten minutes on a Sunday, and that is a cost I am happy to pay, because the review is what makes the content mine.
What doesn't work
I want to be clear about the limitations, because they are the same as in every AI flow I build.
The agent cannot have an experience it hasn't been given. If the editorial plan says "write about invoicing flows" but contains nothing about what an invoicing flow actually looks like in a company with 60 employees, the text turns generic. The quality of the drafts is a direct function of how much concrete material I put into the plan. That is also why the review can't be skipped. It is where my examples come in.
Nor can the agent judge whether a topic is still relevant. If the world has changed since I wrote the topic list, the topic stays on it. So the plan needs reviewing once a quarter, which is a reasonable rhythm for most content flows.
The same setup in an ordinary company
Replace "blog post" with "weekly board report", "summary of customer meetings" or "first draft of a quote" and you have the same flow. A source that people control. A scheduled run. A draft in the right system. A person who approves it.
It is the pattern I build most often, and it works for one reason: nobody has to trust that the AI gets it right. You only have to trust that a person looks before it goes out.
If you have a recurring document that takes half a day to put together today, it is probably a candidate. Describe it in three lines on the home page and I will reply with how I would set it up.
