[overfeed]

> There aren't that many moments in building with LLMs where things line up so the AI can just absolutely nail some code and save me a ton of time.

Your workflow is similar to mine, and not "agentic". The proponents of Agentic workflows would have you handover the bulk of the edits to AI, and you'd hand-hold/course-corrrect its approach via chat while it does the thinking.

I tried the agentic approach for code-gen once, and found it mentally draining. Its like pairing with an over-enthusiastic junior on cocaine that can also type 2000 wpm.

Chunking small changes is great because you don't need the latest and greatest models, Flash variants are more than enough.