Now
The themes I keep coming back to in public work. This is a page about recurring themes, not a private project tracker: it carries no statuses, priorities, or timelines, and it is not a log of day-to-day activity.
AI agents as practical teammates
Not magic, and not a replacement for judgment. The useful version of an AI agent is a teammate with clear boundaries: a defined role, a bounded task, the right tools, and a record of what it did that a human can check. Most of my public writing on agents is about designing those boundaries rather than chasing bigger prompts.
- How I run AI agents explains the split between plan mode and bounded execution, and why a slim harness with narrow roles keeps the work disciplined.
- Day Zero: how my AI projects actually begin is a self-audit of project starts and the ritual that came out of it.
- The layers behind one website revision follows one small change through planning, specialist agents, tests, and review.
Markdown in git and public-safe knowledge systems
Personal and organizational memory should not rot. Plain Markdown files in a git repository are boring in the best way: readable without special software, diffable, reviewable, and easy to check with tests. This site is built that way. Articles live as Markdown with typed metadata, and public-safety rules are checked as contracts rather than trusted to memory.
- Writing is the published output of that system: essays and field notes kept in git.
- The layers behind one website revision shows how frontmatter, canonical links, images, and public-safety rules get checked before a change ships.
- Commitmaxxing and tokenminimizing makes the case for small, durable, understandable steps over heroic final summaries.
Infrastructure small teams can understand
Modern cloud primitives are powerful, and it is easy to bury a small team under them. The theme here is restraint: deployment, observability, and automation that the people running it can explain to each other, with secrets and access handled properly from the start.
- Google Secret Manager: stop hardcoding API keys walks through moving keys out of source code and into versioned, access-controlled storage.
- Google Secret Manager, explained visually is the hands-on companion, with the mental model first and runnable commands after.
- Stats is a small public example of the same restraint: a live activity dashboard that publishes counts only, with no repository names or messages.
How to read this page
Treat this as a standing description of themes, not a changelog. If something here matters to you, the durable version is in Writing; the broader map is on Projects. Anything clinical stays with the practice and off this site.