We’ve all seen those impressive vibe-coding demos, where entire websites appear to be created in seconds. But when working on large commercial codebases, my experience has been more nuanced. Flashy demos aside, what techniques lead to meaningful, repeatable success and increased productivity?
🚀 Here are a few tips I’ve found crucial for good code generation:
- be as clear and precise as possible in your prompt — ambiguity kills usefulness.
- Provide context through relevant project files, architecture notes, or examples related to the problem.
- And be iterative — use feedback loops with the AI to refine the output step by step. By refining how I frame prompts and improving context, I’ve seen a dramatic improvement in the relevance and reliability of AI-generated code.
🐞Beyond code generation, AI has proven incredibly useful for debugging. Feeding in symptoms and providing access to the codebase will allow AI to suggest potential causes, especially for those hard-to-replicate or elusive bugs.
📚 I’ve also found AI responses to be helpful for personal learning — AI acts like an on-demand tutor when working through unfamiliar code, patterns, or frameworks.
⚡ When it comes to improving code quality, it’s a powerful ally for optimisation. Asking how to make your code faster, more resilient, or more idiomatic often yields actionable suggestions.
🔍 It can also assist in Pull Request reviews by catching smaller issues before the human reviewers dig into the bigger picture, like whether requirements have been met or edge cases handled. It can also help generate a summary for a single change and assist with the generation of release notes.
🧪 AI-driven test generation has become another useful tool. It’s a fast way to increase coverage, but it’s essential to remember that those tests reflect current implementations, not the intended behaviour.
This list isn’t exhaustive — new tools and techniques keep emerging, and we’re constantly discovering better ways to work with them.
I'm interested in learning about how you utilise AI. What works? What doesn't?
And no, I’m not naming specific apps, LLMs, or integrations — this is a brief introduction post, and what works for me might not work for you.
Top comments (0)