I doubt this can be changed on a Cursor level due to models so maybe a Composer only change. However, I would absolutely love if when in ask mode it literally asks questions back about uncertain possibilities it can’t confirm.
For example I just had asked Composer to try and narrow down where a bug was likely coming from. It provided 4 solutions after several minutes of thinking. 3 of the solutions were impossible to be the case. If it had just asked the state of things outside of its view/control first instead of just assuming it wouldn’t have wasted tokens and time writing up solutions for the impossible.
I’ve tried to do additional rules and guides to help it along the way I want but with no success.
Hi @ForumDragon Thanks for the post! Whether the agent asks a clarifying question or just runs with an assumption is mostly model behavior rather than a definitive guardrail, but I do have some suggestions for you to try:
Plan mode. This is the built-in clarify-first workflow. Toggle it with Shift+Tab in the editor (or /plan in the CLI). In Plan mode the agent researches first, asks clarifying questions (often as multiple choice), and waits for your answers before proposing an approach, so it’s much less likely to spend time writing up paths that were never possible.
A project rule. Something like “If anything about the current state is uncertain or can’t be verified from the code, ask clarifying questions before proposing solutions” can help nudge this in regular chat and Ask mode too. This is a helpful nudge, but it’s not definitive and can’t be relied upon 100% of the time.
Try a different model. Models vary a lot in how readily they ask questions versus assume. If the one you’re using jumps straight to solutions, switch models in the picker and see which asks the way you want. Some of the frontier models (for example Claude Opus) tend to ask more clarifying questions by default.
Where the agent is being launched from also matters. The Cursor IDE is built for back and forth conversation with the agent whereas Cloud agents are more “set and forget it”. Cloud agents are programmed to work as autonomously as possible with as little input from the user as possible. But it sounds like you’re already in the IDE but just wanted to confirm.
Ty for the response @kevinn! Still looking for more just direct back and forth mode like I’m asking and we have a conversation, not just ask and forget. Best example is if I was talking with a co-worker about an issue as they would know the general codebase already but instead of making assumptions would quickly ask, “did you make sure A,B,C? Yeah. Oh okay then let’s tackle it from another angle. Maybe it’s…”.
Plan mode has the downfall of it creating a plan. It does work great for me and it’s my favorite mode by far. But if I’m trying to find quick solutions to a bug a full on plan isn’t really what I’m looking for. The plan mode these day also create a lot of details, architectural design, and lots of notes about what’s not in scope. None of that is helpful in this situation.
Already have that and tried variations but doesn’t seem to make any difference with or without it.
The frontier models like Claude Opus while do ask more questions they tend to overthink and draw in too much context for simple asks. That’s why I like talking with Composer is it seems to get more to the point quicker.
Pure IDE use, never cloud agents.
Maybe I need to make a “co-worker” mode extension lol