# I wrote a simple project

Solving the problem of having AI review its own code
Currently, most of AI's decisions can already be adopted. Sometimes it asks for opinions just to ask, clearly strongly recommending option A, yet still insists on generating options B and C for me anyway.
The project Orchestrator does this: it breaks down the "one AI handles everything" workflow into four completely independent Claude sessions
CEO handles conversations with me, defines goals and acceptance criteria, and sets the highest priority rules.
Planner receives the goals and breaks down the implementation plan.
Coder writes code according to the plan, raises their hand and says they're stuck if they can't proceed.
Reviewer does read-only review, can only point out problems, can't make changes themselves.
Each session can only see what they're supposed to see. Permission boundaries are clearly defined, and permissions are given to let them iterate on their own. After completion, they report back to me as a whole. Only real decision-making problems come to me.
It's pretty great. After developing more projects, I'll share more insights!