cursor-best-practices

Activate Plan Mode with Shift+Tab so the agent researches your codebase, asks clarifying questions, and creates a detailed implementation plan before writing code. Use it for complex features, architectural changes, and large refactorings, but skip it for quick fixes.

Rules are static, always-on context loaded at the start of every conversation, such as code style and build commands. Skills are dynamic capabilities loaded only when the agent decides they are relevant, keeping the context window clean.

Start a new conversation when moving to a different task, when the agent seems confused or repeats mistakes, or after finishing a logical unit of work. Long conversations accumulate noise that reduces agent effectiveness.

Use your agent harness's parallel execution features, which often create isolated git worktrees per agent. Running the same prompt across multiple models lets you compare solutions side by side and pick the best result for hard problems.

Ask the agent to write tests first and confirm they fail, commit the tests, then instruct it to write implementation code without modifying the tests until all pass. Clear test targets let the agent iterate and self-correct effectively.