With AI, it's best practice to create a plan, thoroughly review it, and then implement that plan.
The plan is the more complex part - similar to [[20260605040356-spec-driven-development]], it is the source of truth for the actions, and thus can even be invoked with a more powerful agent. It's imperative that you actually read over the plan though. This is the source of truth for the execution. Don't just assume it's correct. Similar to [[20260604045417-ai-software-development-loops]]
In plan mode, you iterate with the agent on what you want to build before any code gets written. You discuss requirements, validate assumptions, and refine your approach. The agent explores your codebase and builds context. By the time you start coding, the agent knows exactly what to do and has all the information it needs.
https://www.aihero.dev/plan-mode-introduction
This is what [[20260413103002-ai-compound-engineering]] uses as a framework too. It ensures you're solving the right problem, and in the right way. If the fix is small though, you can usually go straight to implementation