Cursor Prompts

Cursor is a fork of VS Code from Anysphere (founded 2022 by an MIT cohort) built around AI-native editing. GitHub Copilot is a plugin that adds inline completion and a chat panel inside any editor. The structural difference is that Cursor owns the surrounding editor, which lets it ship the Composer Agent (a multi-file agent that plans, edits, runs, and verifies). Copilot is catching up with Copilot Workspace and the May 2025 coding agent, but Cursor still has a measurable head start on multi-file work as of May 2026.

Free is the right tier for trying it out and for editors who only need Tab and a few daily Composer runs. Pro at $20 per month removes the Composer cap, gives access to frontier models (Claude Sonnet 4.5, GPT-5.4, Gemini 2.5 Pro), and is the default for serious daily users (verified on cursor.com/pricing on May 26, 2026). Pro Plus at $60 per month raises the limits for engineers who run Composer on every PR. Business at $40 per seat per month adds team admin and centralized billing. Most individual devs are best on Pro.

Composer is the multi-file agent that went GA on February 26, 2025. It plans, makes edits across files, runs commands (test, build, lint), reads the output, and iterates until the task is done. Use it when the work touches 3 or more files and has a clear acceptance check (a test that should pass, a type-check that should succeed, a build that should ship). Do not use it for one-line edits; Cmd-K is faster. Always run it on a feature branch so you can compare the diff before merging.

Rules live in .cursor/rules/ as MDX files. Every Composer session reads them and treats them as persistent context. They are the single highest-leverage Cursor setting because they survive across sessions and across team members. Keep each rule file focused (one rule file per area: style, db, api, security). Anysphere recommends a few hundred lines per rule file maximum. Past that the rules start contradicting and Composer adherence drops. We have seen the same in our testing.

Model Context Protocol is the Anthropic-designed open spec for letting any AI tool reach outside its sandbox to call external tools. Cursor was the first mainstream editor to ship it on November 14, 2024. The most useful MCP servers for daily work are the GitHub server (read PRs and issues), the Linear server (read and update tickets), the Postgres server pointed at staging (read schemas and run safe queries), and a Sentry server (read recent errors). Install only servers you trust because the agent will call their tools.

Both, with a caveat. Cursor's index runs over the whole repo and Composer can reason across the full tree. The caveat is that very large monorepos (over 1M files) lengthen the indexing time and slow Composer's planner. The fix is to use .cursorignore to exclude vendored code, generated files, and any directory the agent does not need to see. Once the relevant tree is under about 200K files, Composer behaves the same as on a small project.

Not for production code. Composer's auto-accept mode is intentionally opt-in because the agent can write changes that pass the tests but break a downstream expectation the tests do not cover. Treat auto-accept like a code review: skim every diff, accept the small ones, scrutinize the cross-file ones. The five minutes you spend reviewing prevents the half-hour you would spend debugging an autonomous bad commit. We have hit this 4 times in 2 months of daily use.

Two common causes. First, the rule file is over 400 lines and self-contradictory. Cut it to one rule per file and keep each file focused. Second, the user prompt explicitly contradicts a rule. The user message wins. If the rule says "always use Vitest" and the prompt says "write a Jest test for this," Composer uses Jest. We have seen Composer adherence climb from 60% to over 90% just by cleaning the rule files. Adherence is real, the model is sensitive to noise.

Claude Sonnet 4.5 is the strongest default for Composer Agent because of its multi-file planning and tool-use behavior. GPT-5.4 and GPT-5.5 win on debugging from a stack trace and on writing concise tests. Gemini 2.5 Pro wins when the work needs the 1M-token context window (reading a very large file or many files at once). The model picker is per query inside Cursor, so the right answer is mixing them by task instead of locking a single default.

Yes, by default. The selected model provider receives the file content needed to answer the prompt. Cursor offers Privacy Mode that prevents the providers from training on the code, and Cursor itself does not retain code under any tier when Privacy Mode is on. For regulated workloads, Cursor Business and Enterprise add SOC 2 controls and the ability to route to a private model deployment. Read cursor.com/security before sending source code that requires a privacy review.

Using Composer Agent like a chat assistant instead of like an agent. The first user opens Composer and asks "how do I add a rate limiter?" That is an Ask-mode question. Composer is for "add the rate limiter and verify the test passes." The difference is that Composer should be given a job, an acceptance test, and a branch. The agent does its best work when the prompt is closer to a Jira ticket than to a Stack Overflow question. Reframe the prompt and the output improves immediately.

Cursor and Windsurf (Codeium) are both AI-native VS Code forks; the practical difference is that Cursor leads on Composer planning and MCP ecosystem maturity, while Windsurf leads on Cascade's autocomplete strength on long files. v0 (Vercel) is a different product: a web-based AI builder focused on producing React, Next.js, and shadcn components from a prompt. v0 is for prototyping a frontend; Cursor and Windsurf are for editing a real codebase. Pick by job, not by hype.