In this post I’ll share my hands-on experience with Cursor and share some insights from their documentation that might change how you work.
My Experience
I was able to turn a dream project I’ve been putting off for 2 years around in less than 24 hrs!
While Cursor initially felt similar to GitHub Copilot, something about it stood out enough to convince me there’s some unique things happening.
Would I purchase a year subscription? Absolutely. I wouldn’t have said about GitHub Copilot just a few months ago.
Usage Economics
In just 4-5 hours of work, I consumed 117/150 premium model requests — a staggering 24% of the monthly Pro plan allocation (500 requests/month). This surprised me.
I kept “auto-select” enabled all the time (Cursor does this to help you select the best premium model by detecting degraded output performance).
When you exhaust your monthly quota, Cursor moves you to a “slow pool” rather than cutting you off completely (docs). Worth noting if you keep “auto-select” on at all times.
At this rate, we’re talking about the following:
- 120 premium requests/day; or
- 60 premium request/day (with Thinking mode); or
- $100 dollars per month (with Thinking mode)
The lesson for us all:
- The ROI for a very high quality AI coding assistant is a complex calculation, especially inside large orgs.
- I highly recommend starting with an experiment.
- Allocate a $100/dev/month budget.
The Documentation
Project Rules (.mdc)
Rules let you hard-code AI behaviors when:
- Encoding domain-specific knowledge
- Automating repetitive workflows
- Standardizing architecture decisions
Agent Modes
There are three working modes. Agent Mode (default) will automatically search the web when needed. It essentially employs the @Web capability behind the scenes to get up-to-date information (coding discussiong, etc).
Notepads (Beta)
I feel like the idea here is to extend linting configs further.
Notepads extend beyond .cursorrules to define:
- Reusable code patterns
- Project scaffolding rules
- Team consistency standards
- Development guidelines
Think of them as living documentation specifically designed to guide Cursor’s outputs.
Tips
Let’s quickly review some tips!
Tip 1: Cursor Tools and Agent Mode’s Impact on Pricing
Tools and Agent mode go hand-in-hand.
With tools, the LLM is able to loop autonomously while chatting with you, exploring your codebase, browse the web, edit your files, and run terminal commands to complete tasks efficiently. It’s an AI-coder on steroids.
Submitting a request with Agent mode enabled can consume up to 25 of these so called “tool calls” - which in turn will consume exacly one request out of your precious 500 “fast-requests” per month. It’s worth noting that Pro plans also have unlimited “slow-requests” per month (although you’re put on a queue with lower priority if you over-consume things).
Because Agent mode can be quite helpful, but costly, the Agent will ask you to hit Continue if tool calls exceed the 25 limit:
Agents can perform up to 25 tool calls per session. If this limit is reached, you can press “Continue” to allow additional tool calls (each “Continue” counts as one request).
Tip 2: Review the Tooltips and Shortcuts!
This is a suddle UI detail in the Cursor IDE, but it helps tremendously! Every button or interactive component in the chat panel, when hovered, provides details like keyboard shortcut as well as a critical message to help guide you. Some models have a pricing 2X factor. They use tooltips to inform you of that so you don’t have to go read the docs. Don’t miss this!
Tip 3: Read the CHANGELOG and forum.cursor.com!
CHANGELOGS are wonderful: https://www.cursor.com/changelog.
The forum is actually richer in content than Reddit or Hacker News - as far as I noticed - so perhaps prioritize that.
Tip 4: Free Models Don’t Count Against Premium Quota
For simple tasks like testing or understanding the Cursor IDE, or generating documentation, it is suggested toggling to a non-premium model.
You can get a list of which ones are free on Cursor’s available models page.
For example - if o4-mini is selected (free), it will not consume your Premium quota. That means you can have unlimited conversations (and have Agent mode fully enabled). Just please, don’t use it for coding. It wasn’t designed for that.
To toggle across different modes and models on macOS (docs):
- Cmd + . (⌘.) opens mode menu
- Cmd + / (⌘/) loops between AI models
Questions for Further Exploration
- Rule Balance: Can too many rules reduce flexibility? Humans also maintain them, so keep it modular.
- Mode Management: When is toggling between Ask/Edit/Agent modes worth the token savings?
- Optimization Strategy: What’s the right balance of features for maximum ROI?
The key seems to be developing a mental model of which features deliver the most value for your specific workflow — then testing.