Writing code is easier than editing code, for humans and for AI.

The leverage isn't in the generated code. It's in encoding your expertise into a system you can run at scale.

Full write-up: tern.sh/blog/ai-code...

The loop: run a batch → find the pattern in failures → fix the prompt → re-run and measure → repeat.

It feels like programming because it is. You're tightening a spec until the output is indistinguishable from code your team would write.

The counterintuitive part: you don't build trust by reviewing every file. You build it by finding your workflow's failure modes.

If step 2 fails on async tests, that's one bug in your instructions. Fix it, regenerate everything.

Step two: treat your prompt like a program, not a wish.

Migrations need stages: gather context, transform, validate, fix errors. One mega-prompt overloads the AI and makes failures impossible to debug.

Step one: turn the overwhelming problem into a spreadsheet. Find the call sites, tag complexity, note which files don't actually need changes.

You haven't started the migration. You've just gotten smarter about the problem.

This is the central problem of AI migrations: a bot can change 500 files overnight, but you're left with 50,000 lines of potential garbage to review.

Better models don't fix this. The fix is a workflow.