Good point @jeremias.codes! I was assuming teams that already know what debt to tackle but can't prioritize it over features—the classic "we'll fix it next sprint" that never happens. 😅

However, AI development creates new technical debt too. Especially when using multiple agents for feature development, they tend to solve problems through duplication rather than integration—generating similar solutions independently instead of reusing existing patterns, or creating new abstractions

It happened a similar thing to me with Airmail, but I found the experience with Mimestream way superior. Without super fancy features or AI everywhere. It just works.

I've been recommending this pattern to my teammates. It's been helpful for those tedious migration tasks.
Curious if others have found similar patterns when working with AI tools?

It wasn't perfect (multiple bash scripts instead of a unified codemod), but is fascinating that it came up with the pattern by itself.

One interesting observation, while testing Claude Code with a similar migration task, it actually switched strategies mid-way through. After making a few individual changes, it started building migration scripts on its own.

The key insight: for some tasks, asking AI to build tools that help you is often better than asking it to do the task directly.

The remaining 5% were edge cases and some legacy legacy tooltip variants I wasn't aware of. Still, much more efficient than manual updates or having AI attempt each change individually.

This is something I've already used at @buffer.com, we were migrating tooltip components from our old design system to a new one. Rather than having AI update each instance, I showed it one manual migration and had it write a codemod.

~95% of components migrated perfectly in one go.

Instead of asking AI to perform the migration directly, I do this:
- Manually migrate one instance first
- Get the diff of my changes
- Share the old/new component/function signatures with AI
- Ask it to build a codemod that automates the rest