I can't even get LLMs to reliably use tool calls instead of bash, let alone follow existing patterns in a codebase.

What do your prompts look like?

Mine are pretty robust and articulate. I tend to write very lengthy instructions and include snippets of code, file paths, struct names, etc.