What I do
Every piece starts with me. I choose what to write about, usually a problem from client work or something I built, and I decide the angle: what the piece argues, who it is for and where it stands against what else is out there.
The evidence is mine too: the systems I built, the evals I ran, the spend I measured and the client engagements in the case studies.
What Claude does
Claude, Anthropic's AI model, working in Claude Code, turns that brief, my notes, eval results and session records into a first draft. The draft then goes through checks before it can go live:
- Fact-check. A separate agent checks every figure, date, version and named claim against its primary source (the repository, the model card, the documentation, my own logs) and marks each one verified, wrong or unverified. Wrong claims are corrected; unverified ones are sourced or cut.
- Cross-model check. Technical pieces are also read by a model from another lab, OpenAI's Codex, against the code they describe, to catch what a reviewer from the same model family misses.
- AI-tell pass. Stock phrasing and the sentence patterns that mark machine writing are stripped out.
- Editorial review. An editor agent reviews the piece against my tone of voice and has to approve it.
Human in the loop
The checks run on their own; the decisions don't. I stay in the loop at each step that changes what a piece says:
- Before drafting. I set the topic, the angle and the evidence the piece rests on.
- During drafting. When a choice comes up, such as which example to use or how far a claim can go, Claude asks and I decide. I also redirect drafts as they take shape; the guide on local transcription gained its main argument that way.
- Before publishing. I read the finished piece on a private staging copy of the site. When a draft says something I don't hold or can't back, it comes out.
- Going live. Nothing publishes without my explicit confirmation, every time.
I answer for every claim on this site.
Why I say so
I write about putting AI to work. Hiding that I use it for my own writing would undercut the rest of the site. Saying so also shows the method: the model selection guide describes how I route drafting, including why it stays on Opus after Sonnet skipped the editorial review in my drafting eval.
Corrections
If you find an error, tell me through the contact form. I fix it.