I Didn't Have AI Fix the Code — I Redesigned the Structure
At first, whenever a bug appeared, I followed it myself.
Track down where the error came from, instruct the AI to fix that specific part, review the fix, then find the next problem. Working through it one by one, the flow kept breaking. I had to stay inside the debugging structure the entire time. It was doable without knowing how to code — but the effort was heavier than expected.
This time, I tried a different approach.
Separating Judgment from Implementation
Meta-Chulbuji took on the role of seeing the overall flow and priorities. Deciding what needs to be solved first in this system, and which direction makes sense. Actual code analysis, fixes, and verification were handled by a specialized dev AI and Claude Code.
What I did was hold the direction in between.
The result: 11 items were overhauled in two hours yesterday evening. Cash balance field, API retry logic, duplicate execution prevention, RSI method switch, trailing stop introduction — all of it. I don’t think that pace would have been possible if I’d been chasing each bug individually.
What This Work Was Really About
Saying “I had AI fix the code” is only half right.
More precisely, I redesigned the AI collaboration structure. Into a structure where I don’t have to follow everything myself. Into a structure where focusing on direction and judgment makes the rest follow.
My inability to code hasn’t changed. What changed was where I need to focus.
How Operations Will Work Going Forward
This experience established the collaboration structure for the auto-trading project.
With Meta-Chulbuji: assess direction and current state. What condition is the system in, what needs attention next, is the strategy tracking correctly. With the specialized dev AI: handle actual implementation and verification. Code-level work goes there.
With roles divided, where I need to be became clear.
Today I run the first paper trading live verification. The structure has changed — now it’s time to confirm it with results.