Working with LLM-based coding assistants has changed how I write code. Here are some practical tips that have improved my workflow.
1. Speak Your Native Language
This is for anyone who doesn't speak English as their first language: just use your own language.
Your AI agent probably speaks your language better than you do. I see so many developers struggle to express something in English, wasting energy trying to find the right words. Even worse, a lot of your intention gets lost in translation.
In your native language, you can express yourself much better. The AI will understand you perfectly, and you'll communicate your ideas more clearly.
2. Use Your Microphone
Use voice input as much as possible. Yes, it's awkward at first. Your sentences might come out weird, and you'll feel self-conscious.
But here's the thing: when you speak, you naturally give more context. In my experience, even my stumbling and the transcribed "uhhs" provide useful information to the AI. It picks up on the nuances of what you're trying to say.
After the initial awkwardness passes, voice input becomes much faster and more natural than typing.
3. Set Up Global Agent Guidelines
This tip has saved me so much frustration: maintain a global set of rules for your AI agents.
Check out my example: github.com/rensjaspers/agent-guidelines
I use Cursor, and I've set up a global rule that tells the agent to check these guidelines. Here's how this works in practice:
Every time an agent does something annoying, I update my global guidelines file. That problem will never happen again in any of my projects.
I also maintain topic-specific guidelines. In my main file, I have an index section that references nestjs.md, angular.md, and other framework-specific files.
Of course, you can still add project-specific instructions in your project repository when needed.
4. Use the Best Model Available
Unless you're 100% certain that a simpler model can handle the task, just use the best model available.
Cheap models are fine for simple tasks. But as soon as there's any serious logic involved, do yourself a favor and use the best model you can get.
Every single time I tried to save tokens by using a cheaper model for a complex task, I paid for it with my own time. The cheaper models produced nothing useful, and I ended up doing the work over again with the better model anyway.
Your time is more valuable than the cost of using a better model.
Final Thoughts
Experiment with these tips and see what works for you.