How I build with AI

a A workflow that tracks every task

I call it Organic Driven Development (ODD). Every request is authorized, explored and sized before any code changes. Larger work gets a feature document in the repo, odd/tasks/<feature>.md, with numbered specs that quote my request verbatim, a task table and a log. The agent then works task by task, test-first where it applies, commits reviewable work units, and reports what was verified and what was not. Push, pull requests and publishing stay my decision. You can see it in practice in the build session for this site.

b A learning vault

I keep my learning notes in an Obsidian vault, organized per project. Each project runs in one of three modes: learn, log or off. Conceptual questions become concept notes. The vault never copies repo docs, so it stays a record of what I learned, not a second copy of the code.

c Project exams

At milestones the agent writes an open-answer exam about the project and grades my answers. The point is simple: I want to be able to explain the code the agent helped me write.

d Memory

I use Engram, a persistent memory, to save decisions, bug fixes and discoveries across sessions. After a break or a context compaction, the work can resume where it stopped.

I wrote about the learning part in more detail: How I Built a Learning System Inside My AI Coding Workflow (Using Obsidian) on DEV.