Work in progress. My setup changes as the models and tools do, so this page does too. What’s here is how it stands on October 14, 2026.
When people ask how I work with AI, they usually mean three things. Can you set it up, not just use it? Does anything run without you? And how do you know it’s working? Here are my answers, each one linked to the Study that shows the work.
Configure
The setup has five layers: standing rules, skills, connections, checkers and measurement. Each one has a single job.
| Study | What it shows |
|---|---|
| Skills, connections, and the checkers between them | the five layers, how skills and MCP connections are set up and governed, and how the whole thing gets retuned |
| None of my agents can edit a file | four read-only checkers, the moment each one runs, and what each one never gets to see |
| How I work with AI | the working method in one page |
Automate
Routines that run on a schedule, a job that catches up on the ones a sleeping laptop missed, and a packager that won’t ship with a leak.
| Study | What it shows |
|---|---|
| Automation that runs while the lid is closed | nine routines, the catch-up job, and the two rules that keep them honest |
| Directing code I couldn’t write | supervising AI-written code that senior engineers review |
Measure
Claims I can show, and the ones I can’t yet.
| Study | What it shows |
|---|---|
| My setup didn’t make the model better | a benchmark, then the same runs with the setup removed |
| I configured three agents. They sat idle for two months. | a usage audit that overturned one of my own rules |
| Measure it or don’t claim it | why feeling faster isn’t evidence |
| The five rungs | where AI entered one feature, and what each step cost |
What’s still open
Team adoption. I’ve packaged the setup for two teammates and haven’t measured their use yet. When I have those numbers, they’ll get their own Study, and this page will link to it.
If you’re building a team that works this way, let’s talk.