How the loop works

1. 1

   Build the first feature by hand

   No AI on this one. It sets the patterns, the naming, and the structure that everything after it gets measured against. If the foundation is generated, there is nothing to hold the rest of the codebase to.

2. 2

   Generate custom rules from it

   That first feature becomes the source of truth. The rules capture how I structure components, how I name things, how I handle state, and where the edge cases live. They are my patterns, not generic best practices someone else published.

3. 3

   Build the next feature with the rules

   The AI writes it from those rules plus my direction. I read every line before it lands, the same way I would review a teammate's pull request.

4. 4

   Examine, test, and sharpen

   This is the loop. Whatever the model got wrong becomes a new rule, so the next pass starts from a sharper baseline. It usually takes five or six features before the output stops needing correction.

5. 5

   Compose commands from multiple rules

   Once a set of rules holds steady, I bundle them into a single command. That is where the commands below came from: repeated work, collapsed into one instruction I can run on the next project.

6. 6

   The part that stays manual

   Manual testing, code review, and continuous context engineering. AI generates 85 to 90% of the code, but the remaining slice is the part that decides whether the feature actually ships.