Cursor has several ways to steer the agent. They overlap if you are sloppy. They stay out of each other’s way if you pick the narrowest tool.
Rules
Markdown with YAML frontmatter, saved as .mdc under .cursor/rules/.
Use a rule when the instruction should shape many conversations: coding standards, product voice, “never use this API.”
Activation is the whole game:
alwaysApply: true— every chat. Keep these short. Two or three is plenty.globs— when matching files are in context. UI craft belongs here, not in every backend task.descriptionwithout globs — the agent decides if it is relevant.- Manual — only when someone
@mentions it.
A 2,000-line always-on rule is how you get ignored.
Skills
A folder with SKILL.md. The agent sees the name and description, then reads the rest when the task matches — or when you type /.
Use a skill for a procedure: audit this UI, write a skill, brief a cloud agent. Steps, guardrails, definition of done.
If you find yourself pasting the same 20-line brief every Monday, it is a skill.
Prompts
A prompt is a brief you paste. No file required. Promote it when reuse starts.
Commands (legacy slash files) are being folded into skills. New work should be a skill or a prompt, not a new command format.
Tools
MCP connects the agent to external tools. It is not a substitute for the text you put in front of the model. This studio stays on rules, skills, prompts, and recipes.