Cursor AI Chat Shortcuts

Opening Chat (Cmd/Ctrl+L) reveals a sidebar panel functioning similarly to a chat interface, where you can ask questions about how existing code works, request the AI generate a new feature spanning multiple files, or work through debugging a problem conversationally — better suited to open-ended or exploratory requests than the more scoped, immediate nature of inline editing. Adding code to Chat context (Cmd/Ctrl+Shift+L) explicitly attaches your current selection as grounding context for the conversation, ensuring the AI's next response is informed by that specific code rather than requiring you to manually describe or re-paste it — particularly useful when discussing a specific function or file section without needing to retype or explain it from scratch. The @ symbol inside the Chat input goes further, opening a fuzzy picker for explicitly attaching any file, function, or connected documentation source to the conversation regardless of what's currently selected in the editor, giving you precise control over the AI's context rather than relying entirely on its own automatic retrieval of relevant code. Agent mode, toggled with Cmd/Ctrl+I, shifts the interaction model from turn-by-turn conversation to something closer to delegated task execution — the AI can search the codebase on its own initiative, run terminal commands to test its own changes, and work through a multi-step plan across several files, checking back in at defined points rather than after every micro-decision. This is meaningfully different from a standard Chat exchange, where every proposed change waits for your explicit review and approval before anything happens. Starting a new conversation (Cmd/Ctrl+N within the Chat panel) clears previous conversational context, which matters because a long-running chat's accumulated history influences how the AI interprets subsequent messages — starting fresh when switching to an unrelated task avoids the AI drawing on irrelevant prior context from a different part of your work, which can otherwise subtly bias its answers in ways that are hard to notice until a suggestion feels oddly off-topic. A practical distinction worth understanding: Chat conversations in Cursor generally aren't auto-applied to your code the way inline edits directly modify a file — the AI's chat responses may include proposed code snippets or diffs that still require an explicit action to apply them into your actual files, keeping a clear separation between conversational exploration and actual committed file changes, a boundary that Agent mode partially collapses by design once you've opted into its more autonomous behavior. Teams onboarding a new developer onto a Cursor-based workflow generally find it worth explicitly teaching the @ reference habit early, since developers who skip it tend to under-trust the AI's answers simply because the AI was reasoning about the wrong files, not because the underlying model itself was incapable of a good answer. Agent mode's terminal access also means it inherits whatever permissions your own shell session has, so running it inside a project with destructive scripts available (a database reset script, a deploy command) deserves the same caution you'd apply to giving a junior engineer unsupervised terminal access on their first day — reviewing what commands it intends to run, where that's surfaced, is worth doing rather than blindly trusting a fully autonomous run on a sensitive project. A final practical tip for larger teams: because Chat and Agent mode both consume model usage that's typically metered by a subscription plan, teams on a shared or capped plan sometimes establish informal norms around when Agent mode's more expensive autonomous runs are appropriate versus when a quicker, cheaper inline edit or a short Chat exchange gets the job done just as well, treating Agent mode as the tool reached for deliberately for larger, well-scoped tasks rather than the default first move for every change.