Boosting Productivity with Cursor’s New “Plan” Mode

Discussions
1

Hello everyone!

I tried the new feature in Cursor — the “Plan” mode — and I’m genuinely impressed.
My efficiency has at least doubled thanks to it, and I wanted to discuss this with you.

Previously, I had to give very detailed instructions and manually build context for a request to stay efficient and avoid spending time on clarifications or corrections. This is exhausting, easy to overlook details, and the AI can either interpret you too literally or go off track if the instructions are too vague.

With Plan mode, about 80% of these problems are solved. The AI analyzes the code itself and tries to understand exactly what you meant to create a clear and detailed instruction. It doesn’t rush or try to self-correct unnecessarily. When in doubt, it asks clarifying questions. You can also manually edit the plan after it’s generated.

This is very efficient, even if it costs slightly more tokens. You can be sure that you’ll save tokens overall due to the improved efficiency and fewer correction requests. Plus, the plan can be executed with cheaper models.

When the AI follows a detailed plan, it rarely stops at intermediate steps or prematurely claims a task is “perfectly done.” This works especially well with powerful models like GPT-5, Claude Sonnet 4, and Sonnet 4.5.

If you haven’t tried the Plan beta yet, I highly recommend it. If you have, share your impressions, observations, and best practices.

Finally, I want to thank the Cursor team for this update. Yes, it’s a bit late compared to other tools, but in Cursor it’s implemented very well.

I also hope that after Plan mode, Cursor will gain automatic mode switching for the AI, allowing it to work truly autonomously and efficiently.

I also look forward to seeing orchestrator-like features (with sub-agents) in Cursor’s development roadmap.

6 Likes
2

Very glad to hear it!

1 Like
3

Is there a particular model that works best in plan mode? Is auto appropriate here?

4

Yes, Auto mode also recognizes the planning mode.
Most of the time, Auto picks GPT-5 or Sonnet 4.5 for planning tasks.

1 Like
5

I don’t feel its a new thing. If you’re familiar with prompt engineering - doing planning was completely possible in the agent mode. Infact, Custom mode solved this problem for me where I added instructions for the model to do deep research and ask me questions if needed, and generate a report with actionable items in a markdown file.

Plan mode just stole the role of my custom planner mode lol.

But its working fine, I’ve used it the whole day today - and its very model biased. Use the wrong model and it will just ruin your plan.

Recommended to use Claude sonnet 4.5 thinking as they showed in cursor’s youtube video. For me Gemini 2.5 pro worked well with Plan mode. Opus 4.1 was good too, but I lost a whole bunch of requests and tokens because of it - when Gemini did it in one request beautifully.

GPT-5 did not fare well in my testing - the codex and high mode took 10 minutes to create a plan which was either half baked or discussed something else entirely.

6

Yes, of course — that’s exactly how it works.
Previously, you had to switch it manually, and often it just felt like too much effort for minor tasks.

It’s worth noting that the Plan mode also came quite late — other AI coding tools had implemented it long ago.
However, I work with Cursor, not those other tools, and this update has genuinely helped me and sped up my workflow.

It’s implemented much more cleanly and reliably than in other tools (though it still has some bugs). That quality makes a big difference for me.