If language supports comments and LLM is allowed to write them (or docstrings, or any such), there go your tokens. Plus, they will strongly "pull" the context

If language supports comments and LLM is allowed to write them (or docstrings, or any such), there go your tokens. Plus, they will strongly "pull" the context when LLM parses it back, to the point of overriding your instructions (true story)