Hints when agent goes awry, like /btw in Claude

When I’d see the agent/LLM start going into “left field”, I used to be able to add an “echo” with some hint of how it was going astray into commands I needed to allow… and would see it immediately see the echo and say something like “the user says…” and redirect itself accordingly. That’s stopped working, even when I beg it to look for those hints, and it says it will but will not. I see other platforms, like Claude and Codex are allowing /btw or other ways to insert hints while it’s working. Does Cursor have a way of doing this?

Hey there,

Yes – Cursor does have built-in ways to redirect the agent while it’s working. Here’s how:

  1. Immediate messaging (Cmd+Enter): While the agent is running, type your instruction in the chat input and press Cmd+Enter (Mac) or Ctrl+Enter (Windows). This interrupts the agent at its next natural break point and immediately incorporates your new instruction.
  2. Queued messages (Enter): If you press just Enter while the agent is working, your message gets added to a queue and the agent processes it sequentially after finishing the current task. You can queue multiple messages and even drag to reorder them. This is useful when you don’t need the redirect to happen immediately.
  3. Checkpoints for recovery: If the agent has already gone too far off track before you can redirect it, you can use Checkpoints to restore your codebase to a previous state. The agent creates snapshots before making significant changes, so you can click any checkpoint in the chat timeline to preview and restore.

Rules for proactive guidance: If you find yourself frequently needing to redirect the agent on the same topics, you can set up Rules – persistent instructions that guide agent behavior from the start. Great for project-specific conventions or constraints you want the agent to always follow.

The echo workaround you were using was creative, but since it relied on the agent interpreting terminal output as user guidance – which isn’t a guaranteed behavior – it could break across versions. The Cmd+Enter immediate messaging uses the agent’s official input channel, so it’s the reliable way to steer it mid-execution.

More details: Agent Overview - Queued Messages

Best,
Mohit

Each cmd-enter docks me credits :frowning:

Best to just wait it out and see if it figures it out by itself.