How I work with AI

What AI does in my work

Many of my notes, decisions, drafts, and projects live as plain files on my computer, in a vault that Claude Code can access. The AI reads those files and pulls the relevant ones to help me with the task at hand. It organizes, cross-references, transcribes, researches, and drafts against material I gave it. It does the bookkeeping my brain won’t hold anymore.

It does research legwork: finding sources, pulling them in, laying them out for me to judge.

And it does production mechanics: it stages posts into my blog system, edits my videos, keeps my Notion board up to date – and a lot more. A staged post or an edited cut is a draft, nothing more. Between an asset existing and that asset being live there’s work – checking, fixing, deciding – and that work is mine.

What AI never does

It never puts anything live. Nothing goes out under my name that I haven’t verified and made mine – every post and every video passes through my hands last.

It never gets to be my only way into my own knowledge. I keep the files readable, I keep some work fully AI-free, and I make sure I can always find things by hand. The system has to survive the tools changing – or not being available. It’s just files.

And it never interprets sources I’m accountable for. When the claim matters – in client work, in anything scientific – I read the source myself and decide what matters. That’s my job; I don’t rent it out.

How I catch it lying

Nine days into using AI, back in 2023, it invented a scientific reference for me – real author, real journal, completely fake paper. I tell that whole story on the about page. The reflex it left me with is the core of my method:

• Every reference gets checked against the actual source before it’s used. A citation nobody opened is a rumor with formatting.
• If no source exists, the answer is “no source exists.” That’s a result, and it’s allowed to be the result.
• Claims get checked one at a time. A paragraph that “sounds right” has proven nothing.

This isn’t paranoia; it’s the current score. An audit published in The Lancet found fabricated references in roughly 1 in 277 PubMed-indexed papers in early 2026, and the NIH now refuses applications substantially developed by AI. The verification layer isn’t optional for anyone.

Where my data lives

In two places – and if, like me, you have to comply with GDPR, this part is really important.

On my computer: my notes, decisions, drafts, and projects are ordinary files I can open without any AI in the room. Different work lives in different folders, and a session sees only the folder I open for it.

On the AI companies’ servers: the tools I currently use are cloud services from Anthropic and OpenAI. My settings are set so that my work does not train their models. However, every conversation I have with them is transmitted to their servers and stored there under their policies. That fact shapes what I am allowed to put into them.

Client work adds hard obligations: I’m based in Germany, so GDPR applies to everything I do with clients, and I have non-disclosure agreements with most of them. The exact rules I hold are below, under “When I work for clients.”

My own proving grounds give me plenty of opportunities to test and play: this site, a hobby blog, a family member’s wiki. And when the system learns, it learns ways of working – figured out in one corner, carried to everything else I do.

When I work for clients

These rules are older than my AI system. I was using AI chat tools in my work years before I built anything: smoothing text, drafting headlines, doing research legwork. These have been my rules the whole time. I’m writing them down now because I’ve started talking publicly about how I work.

No personal data goes into AI tools. Not my clients’ people, not their customers, not study participants. The individuals I work with appear in my notes by role, “my stakeholder,” never by name.

Each client’s work lives in its own separate working environment, separate from my own brands and from every other client. Client work had its own separate notebooks from my second day with this system; in July 2026 I rebuilt that separation to be absolute. I’m a solo practice; I’m the only person with access.

Nothing about my clients becomes public. No client names, and none of what they work on, ever appears in my public system or this diary. Method crosses over; material never does.

AI output is a draft, nothing more. AI assists in preparing drafts; I finalize every deliverable myself, and I read the sources for anything I’m accountable for. I carry editorial responsibility for my work.

Training is off. Turning off model training is the first thing I do on every AI account I use.

If your project needs more (a data-processing agreement, zero data retention, hosting in Europe), that setup exists through the commercial API route, and I can arrange it at cost.

What I disclose

The words: this site is heavily AI-drafted, and I disclose that once, for the whole site, instead of stamping every post. The method is vibe writing – I tell the story, the AI drafts it, and I fact-check, fix, and cut before anything ships. The ideas are mine and the editing is mine; the polish is the machine’s. I edit bad writing because it’s bad. I don’t reword clean writing to sneak it past an AI detector.

The images: if it’s on this site, it’s real – unless I say otherwise right where it appears. A photograph is a photograph; a screenshot is a screenshot. No AI people – no avatars that look real, no generated photos posing as the real thing. My professional photo was touched up by AI a while back. One of these days I’ll get a new one; until then, that’s the photo.

The tools: every paid tool named on this site, I pay full price for. No affiliate links, no sponsorships. Where I use a tool’s free tier, I say so where I name it. If any of that changes, it’ll say so right where it changes.