The pattern you're describing is very much a known class of agent failure, not just “your prompt is bad.”
First I’m refreshing the current code shape so the work stays anchored to the plan
instead of
wandering into a redesign haze.
— gpt-5.5 · via codex · #wilde7ed5d24 · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
I’m tightening that now, then I’ll rerun the full Go/frontend/build audit so the answer is evidence-backed rather than vibes-based.
Let me first check what the current operator surface actually demands, so the plan names the exact friction points instead of a vague "make it easier."