In early November 2026 I started building a personal working harness. I begin an infrastructure engineering job in December and I want a method before day one, not a pile of habits picked up under deadline.
Three ideas shaped it. AI is stateless, so the codebase must hold no process history. The test case library is the only reliable asset. Conventions slow down entropy.
The harness has three parts. A set of working principles in one file. An export script that pushes that one source into each agent's instruction file, so the rules live in one place and every tool reads the same text. A hand-off command that writes the prompt for the next session, because a session that starts from a file is better than one that starts from memory.
The rules export into Claude Code's user instructions. A /hand-off command writes the next session's starting prompt. Every rule that can be checked by a script is a script, so the agent does not have to remember it and I do not have to police it.
This site is the first thing built with the workflow. The content lives in one file, the tests run against the built site, and a link checker crawls every internal page before I ship.
Next I want to try more tools around the same rules, Orca for parallel agents first, and record what changes in a changelog. Week one is the skeleton. The later weeks are where it becomes a habit.