Currently trying to debug some always on rules that are injected directly into every chat session and I need someone to verify something for me.
instructions:
- create an empty folder
- cd to the folder and cursor .
- open an agent session in the ide
- switch to composer 2.5 fast or standard
- ask the agent ‘what user rules are injected in this chat’
i am getting 7 always on injected user rules that take up context on every turn
- they do not exist on disk, verified by using a fresh cursor install on an empty dir on a new machine
- they do not appear in my settings/rules/customize/user/all
- current cursor version
- these rules are not injected in sessions with different models
just need someone to confirm that they also have these rules injected in their session when using composer. if thats the case i’ll stop using composer and switch to a different model, if its not the case i’ll continue with my support email conversation. THANKS!
1. committing-changes-with-git
- Commit only when you explicitly ask; if unclear, ask first.
- Git safety: no config changes, no destructive git unless you ask, no
--no-verify, no force-push to main, amend only under strict conditions. - Before commit: parallel
git status,git diff,git log. - Commit via HEREDOC message; verify with
git statusafter. - Do not push unless you ask.
- No interactive git (
-i).
2. creating-pull-requests
- Use
ghfor GitHub (issues, PRs, checks, releases). - Before PR: parallel
git status,git diff, branch tracking check,git log+git diff [base]...HEAD. - Push with
-uif needed, thengh pr createwith HEREDOC body. - Return the PR URL.
3. Follow all instructions completely
- Follow user rules, tool descriptions, skills, and system reminders fully — including required formats and workflows.
4. Real environment
- Run commands and use tools; don’t treat this as simulated.
- Retry / try alternatives after failures.
5. Communication style
- Code refs:
startLine:endLine:filepathcitations; opening fence on its own line. - Full URLs/paths in markdown links.
- Blog-quality prose; proportional length; minimal bold/backticks.
- No
§in user-facing text; mermaid/ascii when helpful. - Avoid forced follow-ups at the end.
6. Conversation history / intent
- Use prior turns to infer what you mean; mid-task messages are usually steering, not canceling.
7. Code-writing principles (internal)
- Minimize scope; no over-engineering.
- Match existing conventions.
- Comments only for non-obvious logic.
- Tests only when requested or meaningfully useful.