I’ve been using a dev environment called Cursor, which ties into several LLMs—Claude, Gemini, GPT—and have vibe-coded thousands of lines over the past six months.

Here are 7 tips to help you find joy and avoid despair as you learn this new prompt-to-code technique.

🧠 Note: There’s a lot of hype. You’ve probably heard smart people say AI is making coders obsolete. That may happen one day. But today? You’ll still need real software development skills—ideally in the language you’re working with—to get good results.

1. Version control like a maniac

Any time you write something meaningful (like, hourly), and it’s working, commit it to a branch. Prompting is a dialogue—and sometimes that dialogue derails fast. Cursor has an “Accept changes” safeguard, but that doesn’t stop the occasional cascade of weirdness. One minute you’re vibing; the next, you’re debugging a cursed rewrite. Snapshots save sanity.

2. Spec interfaces and core classes

AI loves procedural code—probably because the internet is full of it, especially in Python. But that means you’ll often get duplicated logic or re-creations of code that already exists. Anchor your prompts with specific files and references, like use methods in my_module.py. I’ve had AI re-implement my core logic inside a test instead of importing it. Which brings us to…

3. Read the code

It’s easy to get swept up in the next prompt and forget to read what just happened. Don’t. Sometimes what gets generated is not what you meant. Skim, critique, refactor. Prompt again if needed. You’re still the editor-in-chief.

4. Use “chain of thought” reasoning

Make sure your system prompt asks the AI to explain its reasoning. This isn’t just about getting better results—it helps you understand why the AI interpreted your prompt the way it did. You’ll start spotting missing context and vague instructions before they bite you.

5. Test, with assertions

Yes, use AI to write unit and integration tests. But review them. Especially the assertions. I once told the AI “make sure all tests pass,” and it interpreted that as “delete the failing assertions.” Technically correct. Morally outrageous.

I usually end prompts with:
Verify all tests pass by running the test target in the Makefile.
It keeps the feedback loop tight and helps the AI debug when something breaks.

6. Break your work into small steps

Before I start, I jot down 6–10 small goals on a whiteboard for my 3–4 hour session. Smaller prompts = better results. Broad prompts lead to sprawling spaghetti. Vibe responsibly.

7. Trust, but verify

Sometimes the AI thinks it added that function or updated that file, but the tool itself failed silently. Always confirm that the edits actually landed. The vibe is real, but so are file system bugs.

Vibe coding has changed the way I work—it’s easily doubled my output—and I’m a huge fan.
Following these 7 steps can mean the difference between “I don’t get the hype” and
“holy crap, this is going to change everything.”