Same here. Step 1 is usually a research doc where I simply describe the task and tell it to research the relevant parts of the codebase. This gets refined to a high-level plan, which gets distilled to a detailed step-by-step implementation plan.

When it comes to the actual implementation I prefer to work through it in small steps, where the AI explains to me exactly what it's about to do and why (and I approve) along the way. This enables me to catch it if it's about to do something I disagree with beforehand. And reduces the time I need to spend reviewing in the end.