I will keep an overview here of the ways in which I use „AI“ in my work. For specific test cases and observations, also refer to the recurring segment in our podcast: https://mit-ki-gemacht-677757124433.europe-west1.run.app/
- Knowledge Management: Obsidian
I’ve been growing my Obsidian vault for many years now. It’s my main notetaking app, my platform for research and writing, and my CRM. I save most of what I read or watch online to my vault using a Readwise connection.
The vault can then serve as a basis for working with AI. For example, I use plugins by Smart Connections to embed my writing and improve semantic search. I can chat with my notes. Or I can give AI systems like Claude Code access to my vault folder and have it help me structure or find things.
When I do research, I often do many deep research passes in parallel, using different tools and comparing web sources to specific search runs over my curated knowledge base.
How I make sure this is safe and sustainable: I keep my own writing, saved articles, and AI outputs strictly separated. I don’t want AI outputs to „contaminate“ my vault, so anything any AI system does has to live in a sealed off folder. For sensitive information, I do a scrubbing pass that anonymizes my notes and removes any PII, locally on my Nvidia Jetson Orin Nano. This ensures that I don’t send anything private to any AI provider cloud – especially third party personal information.
2. Writing
I have three golden rules when writing:
- Always think alone first. Any project, essay or research starts with just my brain and a blank page. This way I don’t risk an AI-first-pass narrowing my possible approaches or ideas. Writing is thinking, so for important work, I always write a full first draft before allowing any AI output to connect with my thoughts.
- Ask AI to ask YOU. Most of my prompting is aimed not at letting the chatbot find a solution for me, but to have it ask me challenging questions that enhance my own thinking and writing. I use the AI system as a machine that prompts me, rather than the other way around.
- Use AI to make connections. This is the strength of modern AI systems: They’re great at managing information. Based on my vault, my AI workflow serves up information that is related to what I’m currently working on. This happens in a side tab while I’m writing. For example, the Smart Connections plugin may be running in the background while I’m preparing a talk on AI governance. As soon as I’ve written a paragraph or two, it will use similarity to find related texts from other sources – e.g. a transcribed Youtube video I watched and saved a few months earlier, or a useful essay from some third party. I can then go deeper myself on those forgotten sources – so the AI system is neither researcher nor writer in that context, but a powerful research assistant.
3. Note intake and transcriptions
I don’t transcribe a lot of meetings, because I find it sufficient to just jot down a few lines to retain the most important info (by hand or by typing into Obsidian). However, when I do think that a meeting will have a lot of dense information and follow-up items, I use my jailbroken Rabbit R1 to record it. As soon as it next connects to my home network, it sends the recording to my Raspberry Pi, which wakes my Orin Nano, which transcribes the meeting using a local Whisper model. A cloud AI then creates a summary and action items (after a local PII-scrub-pass).
I bought the R1 when it came out, just to test it, and didn’t find it very useful. But I like the hardware design, so I now play a lot with custom apps. For example, I’m constantly working on my One-Line-Intake-Tool. Currently, at a double-press of the R1’s action button, it immediately starts the recording-transcribing-scrubbing-summarization workflow. At long-press, I can take a photo (and optionally add a voice recording or typed input), which also gets sent to my Obsidian inbox.
4. App over LLM, flexible APIs over lock-ins
AI moves fast. Model capabilities keep changing, as do guardrails and pricing. So whatever I do, I try to keep the foundation flexible and portable. That’s what I like about Obsidian: All my information lives in markdown, on my own devices. And that’s what I try to achieve with the AI strategy for my clients: How to build AI systems that can be used reliably and sustainably in a constantly changing environment?
Two places to start: Use the AI models you currently love to build robust apps. Don’t rely on their being there forever. Use them to learn and to hardcode something that will work independently of model economics.
Think in connections. Think in interfaces. Where do different qualities of different tools intersect? How does your technology have to work together? How can you make those interfaces robust and flexible? The best technical infrastructures I’ve seen were connections of APIs. With my personal AI use, I try to use APIs as much as possible. Sometimes in connection with local models. APIs to open-source or open-weights models where available and feasible. But in practice, for some tasks you will want frontier models. And not many people have a powerful data center at home, so you need a combination. My advice is just: Whatever you do, know that the AI economy is a beach full of sandcastles. For important workflows, invest in thinking about how to build robust foundations.