It seems my Opus 4.6 as well as other models tend to write their own annoying dark watermarks in the code… so annoying. Like the em-dash, 2-space //, weird separators, and other things even rules don’t work every time… if this is from Model Provider side? Or cursor can prevent it? Do you have similar observations on this?
Hey, this is mostly model-side behavior. Each LLM has its own formatting habits from training, like specific dash choices, comment styles, and separators. Cursor can nudge this with system prompts, but it can’t fully override the model’s default patterns.
A few things that can help:
Project rules with very explicit instructions. Create a .cursor/rules/code-style.mdc with alwaysApply: true and be as specific as possible. For example:
- Never use em-dashes in code or comments, use regular hyphens
- Use standard single-space comments: // comment, not double-space
- Do not add decorative separators or section dividers
- Match the existing code style in the file exactly
Use a formatter like Prettier or ESLint as a safety net. It catches most style issues automatically after generation.
Try different models. Each one has its own fingerprint, and some are cleaner than others depending on the language and codebase.
Rules won’t work 100% of the time since models can still drift, but explicit, specific rules work better than vague ones. A linter plus formatter is the most reliable way to enforce consistent style.
Thanks. Using cursor rule is the best I can do, hope can compete.