YC
How To Get The Most Out Of Vibe Coding | Startup School
@ Y Combinator
04/25/25
Takeaways
TB
Tom Blomfield
04/25/25
@ Y Combinator
Vibe coding is a practice that can significantly improve your coding skills if you're willing to experiment and adopt best practices.
TB
Tom Blomfield
04/25/25
@ Y Combinator
The best techniques in vibe coding are similar to those used by professional software engineers, focusing on achieving the best results with available tools.
TB
Tom Blomfield
04/25/25
@ Y Combinator
When the AI development environment gets stuck, try pasting your code directly into the LLM's website to get a different perspective on the problem.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Using both Cursor and Windsurf simultaneously can enhance your workflow, allowing you to work on front-end and back-end tasks more efficiently.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Think of AI as a new programming language, where you provide detailed context and information to achieve better results in vibe coding.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Start vibe coding by crafting your test cases first, then let the LLM generate the code based on those tests, ensuring quality control.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Spend ample time using LLMs to define the scope and architecture of your project before transitioning to coding tools like Cursor.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Monitor the LLM's responses closely; if it keeps generating incorrect code, reassess the context you're providing to improve its output.
TB
Tom Blomfield
04/25/25
@ Y Combinator
For beginners, tools like Replit or Lovable offer a user-friendly interface to experiment with coding directly, bypassing traditional mock-up designs.
TB
Tom Blomfield
04/25/25
@ Y Combinator
If you have some coding experience, jump straight into tools like Windsurf, Cursor, or Claude Code to start building your projects effectively.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Before writing code, collaborate with the LLM to create a comprehensive project plan, which you can refer back to throughout the development process. After creating the first draft of your plan, remove features you don't like and mark certain features as 'won't do' if they're too complicated, while keeping a section of ideas for later to inform the LLM that they are out of scope for now.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Work with the LLM to implement your plan section by section, explicitly stating which section to work on at a time, and check that it works before moving on.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
I wouldn't expect AI models to one-shot entire products yet, especially if they're complicated. I prefer to implement features piece by piece and ensure each step works before proceeding.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Use version control religiously, like Git, to manage your code. Always start with a clean slate before adding new features to ensure you can revert to a known working version if needed.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Don't hesitate to reset your code if something isn't working. Accumulating layers of bad code from multiple prompts can lead to confusion, so it's better to start fresh with a clean codebase.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Write high-level integration tests instead of low-level unit tests. Simulate user interactions to ensure features work end-to-end before moving on to the next feature.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
LLMs aren't just for coding; they can assist with non-coding tasks as well. For example, I used AI to configure DNS servers and set up hosting, significantly speeding up my progress.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
When encountering a bug, copy and paste the error message directly into the LLM. Often, this is enough for the AI to identify and fix the problem without needing further explanation.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Avoid making multiple attempts at bug fixes without resetting, as this can lead to layers of unnecessary code. Instead, reset and start again to maintain a clean codebase.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
If you're unsure about a bug, try switching models. Different AI models may succeed where others fail, and if you find the source of a bug, reset all changes and give specific instructions to fix it on a clean codebase.
YC
Y Combinator Cast
04/25/25
@ Y Combinator
Write clear instructions for the LLM, using the appropriate naming conventions for each tool. Founders have successfully written extensive instructions to guide their AI coding efforts.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Using local documentation for APIs improves the accuracy of AI implementations. I download the documentation and instruct the AI to read it before proceeding with coding.
TB
Tom Blomfield
04/25/25
@ Y Combinator
The AI can serve as a teacher, helping less experienced developers understand code implementations by explaining them line by line, which is more effective than traditional resources like Stack Overflow.
TB
Tom Blomfield
04/25/25
@ Y Combinator
For complex features, I recommend developing them as standalone projects in a clean codebase to avoid complications from existing projects, which allows for better AI guidance.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Modular architecture with clear API boundaries is beneficial for both human and AI developers, as it simplifies understanding and reduces interdependencies in the codebase.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Choosing a tech stack that has abundant training data, like Ruby on Rails, can enhance AI performance in coding tasks due to its established conventions and similarities across codebases.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Using screenshots in coding agents can effectively demonstrate UI bugs or provide design inspiration, enhancing the development process.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Voice input tools like Aqua can significantly speed up coding by allowing developers to dictate instructions, which the AI can transcribe and process efficiently.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Frequent refactoring is essential once the code is functional and tests are in place, as it helps maintain a clean and manageable codebase.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Experimenting with different AI models is crucial, as their performance can vary significantly across tasks like debugging, planning, and implementing features.
TB
Tom Blomfield
04/25/25
@ Y Combinator
Currently, Gemini excels in whole codebase indexing and planning, while Sonet 3.7 is the best for implementing code changes, although I found GPT 4.1 less effective recently.