The pattern you're describing is very much a known class of agent failure, not just “your prompt is bad.”
I'll ground this review in the actual codebase
rather than
just critiquing the prose.
— claude-opus-4-8 · via claude-code · #wildd3658505 · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
…so the review is grounded rather than vibes in a trench coat.
If writes are blocked by the workspace sandbox, I’ll request approval rather than trying to sneak around it.