Many companies understand the power of how Cursor can increase their developers efficiency and quality through the use of AI, but going in BLIND can cause more problems than benefits.
I have found three different ways over time to get Cursor working with devs and not against them:
- Use Cursor rules for common rules and project-specific rules
- AI model screws up? Add a rule for that.
- Build the intuition on WHICH models are good for WHAT
Table of Contents
What Cursor Rules Are & How to Use Them
The AI world more generically knows Cursor rules as “instructions”. They give direction to the AI models to behave in a way that the developers want. So, if they already tried Cursor and the output frustrated them, they probably didn’t give the AI model a set of guidelines by using Cursor rules.
These two rules help guide the AI model to behave in a very specific way:
Think carefully and only action the specific task I have given you with the most concise and elegant solution that changes as little code as possible.
Do not make any changes, until you have 95% confidence that you know what to build ask me follow up questions until you have that confidence
AI models are like a very smart person with ADHD that’ll think of different possibilities and go through different rabbit holes that you weren’t expecting. Those rules above help prevent most of that by ensuring that the AI model understands the developer’s requirements and preferences with as little assumptions as possible before actually taking action.
Educate Your AI Model Like a Child
Developers will notice over time that most models make the same mistake. Like a child needs reminders such as cleaning up after themselves, AI models need this type of instruction as well.
As mentioned above a common mistake is the AI model going completely off rails making assumptions. Developers may sometimes notice a specific model or group of models behaving in a way that they didn’t expect. They then prompt them in similar ways or almost use the exact same prompt to fix the problem.
If they’ve found themselves in that situation, that’s the perfect time to create a specific Cursor rule to counteract that.
Here’s an example that forces the AI model to stop using NPM, which it typically does, and direct it to use something else:
---
description: Any time any nodejs package management happens to install/update/delete dependencies, run scripts, set configurations, etc
globs:
alwaysApply: false
---
* DO NOT use npm
* ONLY use pnpm
That’s one of many different rules that can be used.
The #1 lesson from this is just to create a habit of creating rules to correct the model and give is necessary context tied to the project. Those rules are versioned within the project, can be shared with other developers, and create a flywheel effect where the model makes less mistakes over time.
An AI model that makes less mistakes opens up more time for developers to spend on important tasks vs fighting the model.
A developer that’s frustrated less is more motivated to use AI assisted coding tools such as Cursor and Claude Code. The boost in moral should directly have an impact in productivity.
Build Intuition on Different Models
A developer that only uses one model is by default not taking advantage of the strengths of others. They’re probably hitting the same walls that one model struggles with over and over.
I suppose it is tempting, if the only tool you have is a hammer, to treat everything as if it were a nail.
Doesn’t make sense for a female to use a hammer to apply eye makeup right? There’s definitely times where a specific AI model just doesn’t make sense to use in Cursor for a particular task.
Here are some general characteristics of a few major SOTA models:
- Gemini 2.5 Pro: Very strong problem solver, able to handle lots of code without getting lost compared to other models, decent at UI/UX
- ChatGPT o3: Very strong problem solver, great at planning, you could think of this one as a product manager, it tends to think a lot, so probably not the best choice in terms of getting a response back quickly
- ChatGPT 4.1: Not much of a thinker as the two above, but can execute on given instructions pretty well.
- Claude Sonnet 4: A very capable model that is very strong at leveraging different tools that are given to it (what makes it more “agentic”) that helps it achieve the task at hand. Best-in-class UI/UX capabilities. Tends to really explore different ideas enthusiastically, which is a strength and a weakness depending on your needs.
- Claude Sonnet 3.5: Similar to ChatGPT 4.1, great at following instructions without deviating
With these tips on how to use Cursor you should see a lot better results. Just remember you can guide the model.
If you liked this article you might also like The AI Design System Workflow.