The 5-step workflow I use to build anything on top of my personal AI system — with a coding agent, without touching a line of code.
Coding agents change how you start a project. The old way: jump into code. The new way: tell the agent specifically what to build and what success looks like, then let it run.
The 5 steps
- Interview — flush out what success looks like
- Spec — user stories of what success looks like
- Build Plan — steps, order, acceptance criteria
- Execute — the agent runs the build plan (Lesson 4)
- Test — conformance checks + run the app (Lesson 4)
+ Loop — when the test reveals something missed, update spec + plan, run again.
Why it matters
Most of the work is upfront. Get the spec and build plan right, and the loop stays small. This is the same flow BrainDrive itself is built on — same 5 steps, same loop, just formalized at scale.
We apply it live to a simple example — chat with documents — to demonstrate. Every step has a prompt file in the template repo at prompts/. Your agent reads them; you direct.
Resources
Full course playlist: Topics tagged paa-course
The Personal AI Architecture: GitHub - Personal-AI-Architecture/the-architecture: MIT licensed user-owned AI runtime: 4 components, 2 APIs, zero lock-in. Your Memory is the platform. · GitHub
TypeScript Template + prompts: ts-architecture-template/prompts at main · Personal-AI-Architecture/ts-architecture-template · GitHub
BrainDrive (MIT-licensed reference implementation): https://braindrive.ai
Discussion
Which step do you find hardest in your own builds — interview, spec, or the loop? Curious where the friction lands.
Previous: Lesson 2 — Free Github Spec & Template for Building Your Own AI. Next: Lesson 4 — Building Private Chat w/ Docs Using Claude Code (Start to Finish).