Below my articles it says that they were written with AI assistance, in the English versions: “This article was translated from German with AI assistance.” What does that mean in practice? Who actually wrote what here? This article shows my path from the first idea to the published article and where the AI does the work and where I do.
The Idea Is Born in the Forest#
I usually start blog articles on a walk through the forest with our dog. The trigger is often a post or a discussion on X, sometimes also a talk or a video on YouTube that I cannot get out of my head.
Often I speak my idea into ChatGPT on my smartphone by voice and let ChatGPT do the research.
Then it continues with a voice chat, a kind of expert discussion, in which I largely set the direction but also let myself be influenced by newly researched information. This way I explore all areas of the topic for the blog article.
I learn something myself along the way. More than once the research has shown that for years I had understood a term differently from how it is meant, most recently with trunk-based development. Then exactly that becomes the topic of the article.
So I do not blog because I know anything better than others, but about my own learning processes.
Sometimes it becomes too much for one blog article. Then I split it up and evaluate the split and the order together with ChatGPT. That is how, for example, one topic turned into a main article with three in-depth articles plus an interactive self-test, which I then published in advance, independently of the series.
At the end I have one or more Markdown files generated as a draft, ideally right away with suggestions for which diagrams would fit.
From Raw Text to Article with a Skill#
I then paste these Markdown files into Claude Code Desktop. Each topic gets its own session and its own branch there.
For the further steps I use my skill blog with these commands:
| Command | What happens |
|---|---|
draft | The raw text becomes the German draft, with a commit. |
review | Optional and repeatable: typos, grammar, sources, consistency and line format are reported, not changed directly. |
diagram | Optional: two to three diagram ideas with placement, then the chosen one is built as an inline SVG. |
translate | English version from the current German state, only on my explicit go. |
preview | The branch is pushed, CI deploys it to the preview. |
publish | The publication date is set to today and the pull request to master is opened. |
xpost | Optional: announcement on X with a diagram as image, posted only on my yes. |
The skill itself emerged from working on an article. After the article Agentic AI Is Like a Rocket I captured the steps that had proven themselves as a skill.
One step is missing from the table on purpose: editing. Editing is not a command, it is my part.
Editing: Sentence by Sentence#
Between draft and publication there are any number of rounds in which I go through the text, first the German version, later also the English one.
Mostly I quote a sentence and write after it how it should read. Often it is just one word, such as a “slightly” before “different focus”, so that the sentence no longer claims more than I can stand behind.
I deliberately do not change the sentence in the file myself. If I give the correction to Claude Code, it may also smooth the phrasing, correct similar passages right away or point out contradictions to me.
Sometimes I just ask what a phrase is supposed to mean. If I do not understand it, the reader probably will not either, and it gets rewritten.
For claims I want to see sources. I have it checked whether the links are reachable and whether they actually fit the passage. If a source is missing, it gets researched; if none can be found, the claim goes.
My articles do not address the reader directly, and they are meant to be written from my perspective. I am not the source of truth, and it should not sound like that anywhere. Often I am only just researching a topic, and then that may show.
The command review only reports its findings.
What gets changed is my decision, preferably as a choice among a few suggestions.
Diagrams#
The first diagrams were Mermaid diagrams directly in the Markdown. I did not find them visually appealing. Now they are inline SVG graphics in the color scheme of my website that work in both light and dark mode.
The AI suggests two to three ideas with placement, I choose, and then it gets built. A diagram is rarely finished at that point. Sometimes an arrow points to the wrong box, sometimes a text sticks out of its frame, sometimes the image says something different from the text. That is part of editing too.
At least one diagram should capture the topic so well that I can attach it to the announcement on X.
Translation#
Translation happens fairly late, once the German text is final. Otherwise every correction would have to be made twice.
The diagrams are translated as well, so that there is a German and an English version of each.
I skim the translation at least once more, too.
Branches, Pull Requests and Preview#
I use branches with pull requests. That gives me a better overview of what is in progress and what I still want to improve or add where, and also regarding the order.
My CI builds a preview with Hugo for every branch. Each branch has its own preview address. It was not like that from the start: At first all branches ended up on the same preview, and the most recently pushed one overwrote the others. With several articles at once, that quickly became annoying.
When an article is finished, the command publish sets the publication date to today and opens the pull request;
I do the merge myself.
After the merge to master, the new article then lands on my website.
I no longer need a draft flag in the article:
Whatever is not finished yet simply stays on its branch.
Ideally Everything in One Day#
It works best when I do everything in one go on a single day. As soon as I do not take the time to finish the article, it usually lies around forever.
There is a second reason: While writing, I either realize that I understand far too little about the topic, or my insight seems trivial to me after the research. That is a trap. What seems trivial to me after the research was not trivial before, and this very learning process is what my articles are about.
And so quite a few articles have piled up, for example on these topics:
- Mutation CRAP: where complexity meets weak tests
- why programming languages could become more important through agentic AI
- developer types in the age of agentic AI
- trunk-based development, features and short branches
- vibe coding vs. vibe engineering
Hopefully I will get around to continuing work on these articles soon.
The Announcement on X#
The announcement posts on X are meanwhile automated as well. When I still did them by hand, I unfortunately often forgot to include an image right away.
The command xpost suggests several text variants to me
and renders the diagram of my choice as an image for the post.
So far, it puts the link to the article into a follow-up post,
because X downranked posts with links.
According to X, that is no longer the case;
adapting the xpost command to that is my next step.
It only posts on my explicit yes.
Conclusion: AI Is a Tool#
Even if the text is generated by the AI, it is strongly guided and edited by me. The idea, the direction of the discussion, the selection and every correction I initiate come from me.
AI is simply a tool, a powerful one at that, but it should not replace people.
Least of all when it comes to visions and ideas.
Sources and Further Reading#
- ChatGPT: research and voice discussion on the go
- Claude Code: draft, review, diagrams, translation and X post
- Hugo: generator for this website and the previews
- Business Today: Elon Musk confirms that X deprioritises posts with links (November 2024)
- Free Press Journal: X Product Head Nikita Bier Confirms Link Penalty Removed Over A Year Ago (July 2026)
This article was translated from German with AI assistance.
