AI summary by TwinMind
How To Get The Most Out Of Vibe Coding | Startup School
Y Combinator
16m 39s
Introduction & Context
- Tom, YC partner, shares month-long experimentation with vibe coding on side projects
- Found vibe coding remarkably effective and improvable through best practices
- Compares current state to prompt engineering evolution 1-2 years ago
- YC Spring Batch recently started, featuring founder tips on AI tools
Founder Tips from YC Batch
- Multi-tool approach: Use both Cursor (faster, front-end/full-stack) and Windsurf (longer thinking time, better for complex tasks)
- Parallel development: Run both tools simultaneously on same project, compare outputs, select best results
- LLM fallback: When AI tools get stuck in loops, go directly to LLM websites (ChatGPT/Claude UI) with same code/question
- Test-driven development: Start with handcrafted test cases (no LLM involvement), use as guard rails for code generation
- Architecture first: Spend significant time in pure LLM defining scope/architecture before using coding tools
- Rabbit hole detection: Monitor for repetitive code regeneration, copy-paste error loops - signals need to step back and reassess
Tool Selection & Getting Started
- Beginners: Start with Replit or Lovable for visual interface, easy UI prototyping
- Product managers/designers going straight to code implementation vs Figma mockups
- Limitation: Backend logic changes unpredictably with UI modifications
- Experienced developers: Jump to Windsurf, Cursor, or Claude Code
- Current preference: Windsurf for complex thinking, Cursor for speed
Planning & Project Structure
- Comprehensive planning: Work with LLM to create detailed markdown plan in project folder
- Iterative development: Implement section-by-section, not one-shot entire products
- Plan management: Delete unwanted features, mark "won't do" items, maintain "ideas for later" section
- Step-by-step execution: Explicitly focus on single sections, test, commit, mark complete
Version Control Best Practices
- Git religiously: Use version control extensively, don't trust tool revert functions
- Clean slate approach: Start new features with clean git state
- Reset strategy: Don't accumulate bad code layers through multiple prompts
- Solution implementation: Take working solution, git reset, implement cleanly on fresh codebase
Testing Strategy
- High-level integration tests: Simulate user clicking through site/app end-to-end
- Avoid unit tests: LLMs default to low-level unit tests, prefer comprehensive feature testing
- Regression protection: Tests catch LLM's tendency to make unnecessary changes to unrelated logic
- Test before refactor: Ensure tests pass before moving to next feature
Non-Coding Applications
- DevOps automation: Claude Sonnet 3.5 configured DNS servers, set up Heroku hosting via CLI
- Design work: ChatGPT created favicon images
- Asset processing: Claude wrote scripts to resize favicons into 6 different sizes/formats
- 10x acceleration: AI as DevOps engineer and designer significantly speeds development
Bug Fixing Methodology
- Error message copy-paste: First step always copy exact error from server logs/browser console
- Error message sufficiency: Often error alone enough for AI to identify and fix problem
- Future automation: Expects tools will soon ingest errors automatically without copy-paste
- Complex bug approach: Ask LLM to think through 3-4 possible causes before coding
- Reset after failures: Git reset after each failed bug fix attempt to avoid code accumulation
- Logging strategy: Add extensive logging when debugging
- Model switching: Try different models (Claude Sonnet 3.5, OpenAI, Gemini) when stuck
- Clean implementation: Once bug source found, reset and give specific fix instructions on clean codebase
LLM Instructions & Configuration
- Custom instruction files: Use cursor rules, windsurf rules, or Claude markdown files
- Extensive instructions: Some founders write hundreds of lines of AI coding agent instructions
- Significant effectiveness boost: Detailed instructions make AI agents "way way way more effective"
Documentation Handling
- Local documentation: Download API docs to subdirectory instead of relying on web access
- MCP servers: Some use Model Context Protocol servers for documentation access (seems overkill)
- Instruction integration: Tell LLM to read local docs before implementation
- Improved accuracy: Local docs access much more accurate than web-based
Learning & Education
- AI as teacher: Use LLM to explain implementations line-by-line
- Technology learning: Better than Stack Overflow for learning new technologies
- Code walkthrough: Effective for understanding unfamiliar coding languages
Complex Feature Development
- Standalone prototyping: Build complex features in clean, separate codebase first
- Reference implementations: Download existing GitHub implementations as templates
- Implementation transfer: Point LLM to working reference, have it reimplement in main codebase
- Modular architecture: Small files and clear modularity benefit both humans and LLMs
- Service-based approach: Shift toward microservices with clear API boundaries
- Consistent external interfaces: Allow internal changes while maintaining API contracts
Technology Stack Considerations
- Ruby on Rails performance: Exceptional AI performance due to 20-year-old framework with established conventions
- Training data quality: Rails has consistent, high-quality training data from similar codebases
- Newer languages: Friends report less success with Rust, Elixir due to limited training data
- Convention importance: Well-established patterns improve AI code generation
Advanced Interaction Methods
- Screenshots: Copy-paste screenshots to demonstrate UI bugs or design inspiration
- Voice input: Aqua (YC company) enables 140 words/minute voice input vs typing
- Grammar tolerance: AI handles minor transcription errors well
- Tool flexibility: Works across Windsurf, Claude Code, and other platforms
Refactoring & Code Quality
- Frequent refactoring: Refactor with confidence when tests are implemented
- LLM-assisted refactoring: Ask AI to identify repetitive code or refactoring candidates
- Professional standards: Keep files small and modular, avoid thousand-line files
- Test-driven refactoring: Use tests to catch regressions during refactoring
Model Performance Comparison
- Gemini: Best for whole codebase indexing and implementation planning
- Claude Sonnet 3.5: Leading contender for actual code implementation
- GPT-4.1: Currently underwhelming - too many questions, frequent implementation errors
- Rapid evolution: State-of-the-art changes weekly, requires constant experimentation
- Specialized strengths: Different models excel at debugging, planning, implementation, refactoring
