Cursor is bigger than a chat box. If every problem is “write a better prompt,” you will fight the product. If every problem is “add another tool,” you will ignore the files that change the next edit.
In the editor
- Rules (
.cursor/rules/*.mdc) — constraints the agent should already know: voice, stack, visual craft. - Skills (
.cursor/skills/*/SKILL.md) — procedures for a job you repeat. - Prompts — paste once. Promote when reuse starts.
- MCP — tools: GitHub, databases, browsers. Useful, and a different job from taste.
- Plugins — packaged MCP plus rules plus skills. Install them when you need the tools, not as a substitute for a short rule.
Around the editor
- Cloud agents — run a slice on a remote machine. Brief them with a prompt or skill; do not ask them to “build a platform.”
- Start from scratch — no GitHub required. Prompt, preview the forwarded URL, then Create repo on Origin when the slice is worth keeping. Vercel publish is a later click.
- Custom Modes — pin a skill for the whole chat (⌥⏎ from
/). Playbooks belong here; taste rules do not. - Subscriptions — wake a cloud agent on a PR, Slack, Linear, or a timer. Write a stop condition.
- `/goal` — a long-lived objective until it is observably done.
- Projects — a coordinator with shared context for a feature or migration that lasts. Plan and delegate; do not treat it as a bigger prompt.
- Self-hosted machines — tool execution on your network (My Machines or a team pool). Inference still runs in Cursor. Default stays Cursor-hosted.
- Rollouts and Security Review — Teams/Enterprise bots. We do not wrap their dashboards. Complementary files: Security PR review, No secrets in git, Make this change observable.
- Models — Grok 4.7 is a long-run picker choice. See Grok long run. Other chatbots are out of scope unless they change a file in
.cursor/.
A simple habit
- Install a taste pack (Studio stack is the default).
- Add only the tools you will actually call this week.
- Delete rules that have not changed an edit in a month.