In general I find it better to use AI to write tools, rather than using AI to replace a tool. AI is [[20260326100028-ai-non-deterministic]], and there's so much nuance in how you set up your context, what model you're using, what harness you're using, etc. You cannot guarantee that the same input will get you the same output.
Instead I prefer to use AI to write specialised tools. AI is really good at making greenfield things, especially when they only need to be throwaway, or are used for internal tooling [[20260601032608-vibe-coding-dangers]]. Instead of many people repeating the same instructions to get similar outcomes, channel your energy into a shared tool.
It's more efficient, it's more consistent, you can catch edge cases, you can improve over time, you can work as a team, etc. It's just the approach of [[20200307105301-dry]], which is good engineering practices.