Write code like a human will maintain it
Points and comments are a snapshot, not live.
Duplicated code trains LLMs to write more of the same bad patterns.
The author notices that letting LLMs copy-paste the same conditional across files-instead of extracting a shared helper-signals a pattern to the model. Each shortcut becomes a signal that the bad pattern is the codebase's style. When the LLM later generates code, it copies the existing duplicated logic instead of deriving a clean solution. The author warns that code smells stack up, making it harder to prompt your way out of a tangled codebase. The lesson: write code as if a human will maintain it, because the LLM will imitate whatever it sees.
What commenters are saying
Top commenters agree with the post, with one calling out the broader principle: 'u right.' A large subthread revives the classic quote about coding for a homicidal maintainer, then pivots to a strong consensus that outdated or verbose LLM-generated comments are more dangerous than no comments. Several commenters share concrete frustration and coping strategies: global CLAUDE.md rules often fail because the model follows surrounding comment style; one user now runs a 50-item automated checklist across 20 subagents to strip 'explain-the-code' comments, vague terminology, and plan-document references from generated code.