Articles
Everything I've written, grouped by what it's about.
Norns and the agents in my terminal
How I work with Claude Code and Codex, and the tools I built to keep track of them.
- The mental load of AI agents after six months with Norns CompanionNorns Companion shows me which agents are waiting for me. The fatigue stayed anyway, and it sits in the decisions, the code review and everything I have to read.
- Claude Norns: experts within reachClaude Norns makes Claude Code feel less like one broad AI and more like a small team of specialists for review, security and architecture.
- Tokens do not measure the work, they measure the uncertaintyThe model that did the most work used fewer tokens. That says more about the instructions than the work.
- Claude vs Codex: same upgrade, two very different estimatesI let Claude and Codex estimate the same Drupal 10 to 11 upgrade using the same skill. The estimates differed considerably, but the interesting part is why.
- How I run multiple AI agents in parallel without losing trackSeveral AI code sessions running at once sounds efficient, until you lose track of who is waiting for you. So I built Norns Companion.
Freya and local AI at home
The family assistant, running locally on a graphics card instead of in the cloud.
- From 25 seconds to 1.8: everything was local, and still hard to useEvery part of my local voice assistant was fast and the assistant was still slow. I trace the latency, split the system into routes and explain the move to Swedish.
- How far does 16 GB of VRAM get you? Gemma 4 12B on an RTX 5060 TiI test Gemma 4 12B as the local brain for my family assistant on a 16 GB RTX 5060 Ti, including performance, failures and the boundary with cloud AI.
Web
Web platforms, which is the day job.