How I Operate

Numbers as of October 6, 2026

I use AI every day as a working partner, and I measure it like a product: I try to log every error I catch, then classify it and audit it against the actual record. The short version: AI is a force multiplier, but it doesn't replace critical thinking, and the receipts prove it.

The system: context is the product

Since May 2026 I've run my job search, my newsletter (The Authenticity Spectrum) and my writing through Claude. The model is frozen between releases, so the learning layer has to live with me. I built it as a set of working files the model reads at the start of every session:

• A standing-rules file: how I work, what's confidential, what's been decided, and where everything lives.
• A voice constitution, built from my logged before-and-after rewrites of Claude's drafts, so drafts sound like me.
• A frameworks corpus: my own stories, positions and frameworks, each with its lineage, so the model argues my point instead of a generic one.
• Repeatable skills for the jobs I do weekly: committing a session's results to the right files, a pre-publish fact check, a four-persona review I call the wisdom council, LinkedIn analytics reads, and a scheduled audit for anything that fell through the cracks.

The operating rule I landed on: patch and feedback, not theme and regenerate. I hand over my own words and the model tightens, checks and reacts. When I let it write from my themes, accuracy and authenticity both dropped.

Grading the AI, then grading the grader

Since August I've kept an oversight log of every time I had to catch, correct or override Claude. In October I asked a harder question: how complete is my own log? I directed Claude through an audit of the full record: 205 sessions, 4,416 of my messages, about 314 hours of active work. Twelve AI labeling agents ran in parallel against the same written spec, a separate agent spot-checked them, and a thirteenth matched transcript errors to my log in both directions.

What the data taught me

• The answer is the product. QA checks how an answer arrives. Evals check whether it's right, and that's now where the value and the risk live.
• Evaluate the evaluator. A golden set built using the tool meant to be evaluated ends up being a self-fulfilling prophecy. My own log and the transcript audit each caught errors the other missed.
• Rank autonomy by blast radius, not visibility. Agents run what can be undone; people own what can't.
• A rule not enforced is just a suggestion. At least 38% of errors broke a rule that existed.
• Who's prompting who. Once I told the model to take what I have instead of reinventing it, I stopped getting a hybrid of me and started getting me, tightened.

AI makes a good team faster. It doesn't decide what good is. That part is still the job.