The call
Check Project Rules or AGENTS.md into each repository, plan broad edits in Plan Mode before the agent writes, and default to Cursor's own models for routine work.
Have the developer read every diff before a second reviewer sees it.
Cursor's agent follows your conventions only when they sit in rules it reads on every request, and each request's cost depends on which model pool it draws from. Settle both before work starts, then keep every agent change small enough for a person to read before it merges.
When you need this
- Engineers get different results from the same request because nothing tells the agent your conventions.
- Included usage runs out mid-month because every request goes to the most expensive model.
- Agent changes reach review larger than a reviewer can check in one sitting.
- Reviewers keep correcting the same mistakes, such as ignored utility functions or outdated framework patterns.
Worked example: rules that turned wrong SQL into usable queries
The task
Salesforce's mobile performance team used Cursor to migrate about 240 dashboard queries.
The first attempt
Without guidance the agent wrote wrong SQL for the team's schema.
What changed
The team wrote explicit rules and a schema mapping document for Cursor to read on every request.
The result
Engineering time for the migration fell 75% by the team's own measure once the rules were in place.
Evaluation criteria
Starter project rules
Does the agent read your conventions before every task?
Look forCheck Project Rules in .cursor/rules or AGENTS.md into the repository. Add Team Rules from the dashboard on Team and Enterprise plans. Start with the corrections reviewers repeat, such as test structure and existing utilities. One Hacker News commenter asks the agent to write a rule after each style correction.
Model-choice notes
Who decides which model handles a request?
Look forThe Grok and Composer models draw from the weekly pool included with each plan. Third-party models bill at their API list price from prepaid credit, from the first request; the Other Models pool exists only on legacy plans. On Teams and Enterprise, Auto with Cursor Router picks a model per request. Engineers report planning on a frontier model and implementing on Composer. Others report Auto sending hard tasks to a weaker model, so write down when engineers should override it.
Plan before larger changes
Does the agent research and propose before it edits?
Look forUse Plan Mode to research the codebase and answer clarifying questions before edits begin, then review the written plan. A team running Cursor on a large monorepo plans first and has the agent implement against tests before anyone reviews.
Keep changes reviewable
Is each agent change small enough for a human to check properly?
Look forKeep one ticket per pull request and give Cursor tests it can run during the task. Have the developer read the diff in the editor before a second reviewer sees it. Agent Review checks local changes after each commit using the same BUGBOT.md rules as Bugbot.
Review checklist
What does the reviewer confirm before approving an agent diff?
Look forCheck that the diff uses existing utilities, follows the repository rules, includes tests that fail without the fix, and touches only required files. Cursor saves agent edits directly to disk. Commit before each run and compare the diff with that commit.
Questions to ask vendors
- Which models on our plan draw from the weekly pool, and which bill at API list price from prepaid credit?
- Which models can Cursor Router choose for our team? On Enterprise it respects model access controls, so who blocks the models we have not approved?
- Do enforced Team Rules reach the CLI and cloud agents? Cursor's Rules docs apply Project Rules the same way in the editor, the CLI, and cloud agents, but describe Team Rules only for Agent chat, so get Cursor's answer in writing.
- Which models need provider data retention, and who on our team approves them?
Red flags
- A rules file copied wholesale from an old style guide, so the agent reads pages it does not need.
- Every request sent to the most expensive model by default.
- Agent diffs approved because the agent's own new tests pass.
- Agent edits that were accepted without anyone reading them.
Products in this piece
Sources
- Official documentation
- 6
- Public source
- 5
- Cursor docs: Rules
- Cursor docs: Models & Pricing
- Cursor docs: Cursor Router
- Cursor docs: Plan Mode
- Cursor docs: Agent Review
- Cursor docs: Agent Security
- Hacker News: having the agent write a rule after each correction (Jun 5, 2025)
- Salesforce Engineering: Cursor cut dashboard migration time 75% across 240 queries (Nov 5, 2025)
- Cursor forum: running Plan Mode, tests and layered review on a large monorepo (Feb 18, 2026)
- Hacker News: Auto routing hard tasks to Composer (Jun 16, 2026)
- Hacker News: reviewing a Cursor-written feature that ignored the utility library (Oct 9, 2025)