I recommend the open-source project plan-tree from this group member. Its approach is to organize all your discussions with AI, including solutions, decisions made, current progress, and where to continue next, into a Markdown planning tree. It doesn't make decisions for you, but rather makes your decisions clearer and more complete. It solves two of the most important things in AI programming. First, it aligns costs. Each time you start a new session, you directly input the planning tree without needing to explain the background again. Second, it forces you to truly think things through. When you're writing your decisions into the planning tree, you'll find that many "almost understand" statements are actually "unclear." It forces you to transform vague understanding into structured text; this process itself clarifies your thinking. plan-tree uses a "large loop" workflow: first, thoroughly discuss, clarify, and document feasible solutions; then, let the AI execute them in batches; after execution, write the progress, evidence, and remaining issues back into the planning tree. Plant the tree first, then let the machine chop the wood. I think this is the proper approach to AI programming. Think about it: the scarcest resource in the AI era isn't execution ability—AI provides that—but rather a sense of direction—the kind of direction where you know "what I'm doing, why I'm doing it, and where I'm going next." A plan-tree is a tool that helps you maintain this clarity. It can reduce deviations in AI's functional implementation. If you have similar issues, I recommend giving it a try.