The LLM is trained on the past. The capabilities have leapt forward, but its stance is pre-AI. Probably by definition, since that is the bulk of its training.

It frets about changing code. It prefers incorporating workarounds rather than recognizing the need to refactor things. Always minimizing the "blast radius".

Behaves like a junior engineer that is hesitant to take on responsibilities.

"Surfaces" and "Flags" things without giving the context, laying things out, but not zooming in on what's relevant, leaving it open ended without a call to action, or a call to action without substance, garnished behind convoluted AI-speak. Like a junior who doesn't know how to communicate into a team.

Endlessly buries issues that are trivial one-liner fixes as deferrals that "land" in session remarks, PR comments. Prefers to pile up technical debt than to fix it and move on. Probably stems from being trained as a junior that must not extend the scope of any task.

But wraps everything in an endless, complex, dense prose that you are constantly having to sift through and try to understand. Bloats code, comments, and documentation.

It's been said already a year ago that AI coding agents are like genius idiots. With each release of an Anthropic model, its personality doesn't really change, it just gets better at producing content.

No amount of CLAUDE.md and rules tuning changed that.

Is there a way?