This content originally appeared on DEV Community and was authored by interviewassistant assistant
Over the past year, we’ve witnessed a transformative shift in software development: “vibe coding” — the practice of using AI tools to generate, debug, and optimize code — has moved from novelty to necessity. As Tom, a partner at Y Combinator, shares from his recent experiments, these tools have become “remarkably good,” and developers can get “measurably better” results by adopting certain techniques.
This evolution echoes the prompt engineering boom we saw a year or two ago, when new techniques emerged weekly and spread rapidly across social media. What’s particularly interesting is that the most effective vibe coding approaches mirror traditional software engineering best practices, suggesting that AI isn’t replacing good engineering so much as augmenting it.
Let’s explore practical advice from YC partners and founders on how to get the most out of these powerful new AI coding assistants.
Starting Your Vibe Coding Journey
If you’re completely new to coding, visual interface tools like Replit or Lovable provide excellent entry points. Many product managers and designers now skip mockups entirely, going straight to code implementation because it’s become so accessible. While these tools excel at UI changes, they can struggle with precise backend modifications.
For those with some coding experience (even if rusty), more sophisticated tools like Windsurf, Cursor, or Claude Code offer deeper capabilities.
Best Practices for Effective Vibe Coding
- Always Start with a Comprehensive Plan Rather than diving straight into code, work with your AI to develop a detailed project plan:
Create a markdown file in your project folder to outline the entire implementation
Go through this plan deliberately, removing anything too complex
Maintain a “future ideas” section for features out of current scope
Implement section by section, testing and committing as you go
Have the AI mark completed sections as you progress
As Tom notes, “I probably wouldn’t expect the models to oneshot entire products yet, especially if they’re complicated.” Breaking projects into manageable chunks ensures more consistent results, though this advice may change as models continue to improve rapidly.
- Use Version Control Religiously Version control isn’t just a good practice — it’s essential when vibe coding:
Make clean git commits before starting each new feature
Don’t hesitate to use git reset –hard HEAD if the AI goes off track
Be wary of multiple fix attempts that accumulate “layers of bad code”
If you finally solve an issue after multiple attempts, reset and implement the clean solution on a fresh codebase
- Prioritize Testing Tests provide guardrails that prevent AI coding assistants from making unexpected changes:
Focus on high-level integration tests rather than unit tests
Simulate real user interactions rather than just testing functions
Ensure tests catch when the AI makes unnecessary changes to unrelated logic
Write tests before moving to the next feature
- Leverage AI Beyond Coding These tools can handle many adjacent tasks that previously required specialized knowledge:
DNS configuration
Server setup and hosting
Image creation and manipulation
DevOps tasks
As Tom experienced, having Claude 3.7 Sonnet configure DNS servers (a task he “always hated”) and set up Heroku hosting “accelerated my progress like 10x.”
- Streamline Bug Fixing When encountering bugs:
Simply copy-paste error messages directly to the AI
For complex bugs, ask the AI to consider 3–4 possible causes before writing code
Reset to a clean state after failed fix attempts
Add logging liberally
Switch models if you’re stuck — different AIs excel at different problems
- Create Custom Instructions Most AI coding tools allow for custom instructions:
Write detailed guidelines for how the AI should interact with your codebase
Some founders maintain hundreds of lines of instructions
These dramatically improve the AI’s effectiveness with your specific project
- Make Documentation Accessible While AI models’ ability to access online documentation remains inconsistent:
Download the relevant API documentation to your project folder
Direct the AI to read this documentation before implementing features
Use the AI as a teacher to explain unfamiliar code line by line
- Handle Complex Features Methodically For sophisticated functionality:
Build a standalone reference implementation in a clean codebase first
Direct your AI to follow this implementation while integrating into your main project
Maintain modular architecture with clear API boundaries
Keep files small and focused
- Choose Stack Wisely Some technology stacks work better with AI than others:
Mature frameworks with established conventions (like Ruby on Rails) often produce better results
Newer languages with less training data available (like Rust or Elixir) may be more challenging
This gap may close as more training examples become available
- Utilize Visual and Voice Interfaces Modern AI coding assistants offer multimodal capabilities:
Use screenshots to demonstrate bugs or design inspiration
Try voice input tools like Aqua to dictate at 140 words per minute
AI’s tolerance for minor errors makes voice input highly effective
- Refactor Frequently Once you have working code with tests in place:
Ask the AI to identify refactoring opportunities
Keep files small and modular
Follow professional software development best practices
- Keep Experimenting The landscape evolves weekly:
Try each new model release
Different models excel at different tasks
For example, Gemini currently shines at codebase indexing and planning, while Claude 3.7 Sonnet often implements changes more effectively
Practical Tips from YC Founders
YC founders shared several practical insights from their daily work with AI coding tools:
“If you get stuck with one AI tool, try pasting your code into the LLM’s direct interface” — Sometimes the IDE integration limits the model’s capabilities
“Load up multiple AI tools on the same project” — Different tools offer complementary strengths
“Think of AI as a new programming language where you program with natural language” — Provide detailed context for best results
“Start with handcrafted test cases before generating code” — Set clear guardrails for the AI
“Spend an unreasonable amount of time planning architecture before coding” — Don’t let the AI freestyle in your codebase
“Watch for rabbit holes” — If the AI keeps regenerating similar code without progress, take a step back
The Future of Vibe Coding
As these tools continue to evolve at breakneck speed, we’re likely to see:
Automatic error detection without human copy-pasting
More sophisticated architecture patterns optimized for AI collaboration
Continued convergence between traditional software engineering best practices and AI-assisted development
The key insight is that vibe coding doesn’t replace good software engineering — it amplifies it. By following these practices, you can harness the incredible productivity of AI coding assistants while maintaining code quality and avoiding common pitfalls.
What techniques have you discovered for effective vibe coding? Share your experiences in the comments!
This article was created based on insights from Tom, a Y Combinator partner, and various YC founders discussing their experiences with AI-assisted coding.
Here is the source of the article: https://www.youtube.com/watch?v=BJjsfNO5JTo
This content originally appeared on DEV Community and was authored by interviewassistant assistant