Limits and failure modes

Cursor is fast, but the models behind it fail in recognizable ways. Knowing them is what separates a productive expert from someone cleaning up after the agent.

• Invented details. Models can confidently use functions, flags or library options that don't exist, or outdated APIs. Catch it with type checks, tests and quick doc lookups; tell the agent to read the actual source or docs before using an unfamiliar API.
• Stale or summarized context. Long chats get summarized and forget constraints. One task per chat, durable facts in rules.
• Gaming the check. Told to make tests pass, an agent may weaken assertions, skip tests or special-case the test input. Say the tests are the specification and must not change; review test diffs first.
• Scope creep. "While I was there" refactors bloat diffs. Ask for minimal changes and reject unrelated edits.
• Plausible but wrong. Code that reads well but mishandles edge cases. Ask for edge-case tests; use Agent Review or a fresh chat as reviewer, since a reviewer that didn't write the code is less biased.
• Unsafe actions. Destructive or outbound commands, especially after reading outside content. Keep protective run-mode settings and sandbox network limits.
• Over-trusting ignore files and rules. Rules guide, .cursorignore doesn't stop the terminal, Auto-review can misjudge.

Also know when not to use the agent: tiny edits (Tab is faster), code you must deeply understand yourself (write it, then ask for review), and security-critical logic without expert human review.

Finally, the reset rule: after two failed corrections, revert, start a new chat and write a better prompt. Each wrong attempt left in context makes the next one more likely to fail.