Build the Expertise Into Each Folder
How I scope AI coding agents to the filesystem: each folder carries its own memory, skills, and conventions, many agents write to one shared log, and the whole setup travels to whoever opens it.
Working with the agent
Agentic Practice
How I configure and work with an AI coding agent day to day: not the systems I build with it, but the practice itself. What I delegate, how corrections become configuration, and why a well-kept companion compounds.
Most of the agentic story on this site is about what I built: the platform, the systems. This is the other half: how I work. The shift isn't only in the software I ship; it's in the day-to-day itself. Configured and refined with some care, an AI agent stops being a tool you reach for and becomes a companion that accumulates, one that remembers your corrections, runs your recurring work, and gets a little more useful every week. The build methodology has its own page; this is its day-to-day counterpart. These are notes on the practice.
How I work with it
- Activation energy is the real costThe hardest moment in solo work is the blank file. The agent makes starting cheap. I get a first pass and edit instead of stare. That alone is the difference between a task shipping this week or living in my head for a month.
- Delegate the recurring, keep the judgmentRepeatable production (boilerplate, document generation, turning hand-edited artifacts into engine-built ones) goes to the agent. What is worth building at all, and whether a confident answer is actually right, stays mine.
- A correction should only happen onceWhen the agent gets something wrong, the fix is not just a one-off edit. I encode the correction (a saved memory, a standing rule, a guard in the code) so the same mistake does not come back. The setup learns.
- The companion compoundsSkills, memories, and small automations accumulate. Each task starts further along than the last because the last one left something behind. The configuration itself is something you refine, not just use.
The starting point is the longer reflection on what an agent actually changes for a solo builder: Building Solo, With an Agent. Everything below is the running, more specific version of that.
Notes on the practice
Teaching It Once
An agent gets better not from cleverer prompts but from corrections that stick: turn every fix into configuration (memories, rules, small skills) so the same mistake never comes back.
Building Solo, With an Agent
I build most of this alone, but not really alone. What working with an AI coding agent actually changes, where it carries real weight, and where you still have to be the one who understands.