What is the correct way to vibe-code Machine Learning projects?

I'm currently learning Machine Learning through a course, and I want to start building projects alongside it.

My main goal right now is simply to build several good ML projects and get familiar with the complete project development process.

I want to use AI coding tools such as Cursor, Claude Code, or GitHub Copilot to speed up development, but I'm unsure about the right way to vibe-code an ML project.

For example, should I:

- Give the AI the complete project requirements and let it build the project?

- First create the architecture/pipeline myself and then let AI implement it?

- Build the project step-by-step and ask AI to implement each stage?

- Let AI handle things like data cleaning, EDA, preprocessing, and boilerplate while I focus on the ML decisions?

- Give AI a detailed specification before starting?

- Ask AI to review and improve the code after it generates it?

- Use one long conversation/context for the entire project, or separate prompts for different stages?

- How should I handle debugging and modifying AI-generated ML code?

Basically, what is the best workflow for vibe-coding an ML project from start to finish?

I'm not trying to replace learning ML with AI — I'm already studying the concepts separately. I just want to use AI effectively to build projects faster without ending up with a messy or poorly structured project.

I'd especially like to hear from people who have built ML projects using Cursor/Claude Code/Copilot:

What workflow do you personally follow, and what mistakes should I avoid?

Also, please suggest any good communities where I can see how other people are building ML projects and discuss AI-assisted development.