Tutorial, Other
How To Get The Most Out Of Vibe Coding | Startup School
- Vibe coding capabilities are expected to improve measurably over time as best practices emerge, potentially reaching the rapid discovery pace of the 2022–2023 prompt engineering era within one to two years, with advice expected to evolve significantly within the next two to three months.
- Tool selection is predicted to depend on experience level: beginners are advised to start with Replit or Lovable, while experienced users can transition to Windsurf, Cursor, or Claude Code, with specific expectations that Lovable struggles with back-end logic while Windsurf is suited for longer thinking processes and Cursor for faster front-end updates.
- A workflow involving the reverse coding of handcrafted test cases is planned to establish guardrails, aiming to eliminate the need for micromanagement once tests pass, with a preference for high-level integration tests over the LLM's tendency to generate low-level unit tests.
- Strategic planning includes spending time defining scope and architecture with an LLM, creating a markdown file to mark features as out-of-scope, implementing projects section-by-section, and committing to Git before moving forward to prevent accidental accumulation of bad code.
- Risk mitigation strategies involve using Git religiously for clean slates and rollbacks, avoiding the accumulation of layered code from repeated prompts, and employing a clean code base reset after failed bug fixes to address issues like non-functional code generation or rabbit holes.
- LLMs are anticipated to act as designers, DevOps engineers, and teachers for non-coding tasks and language guidance, with expectations that future tools will ingest errors directly, tail logs, or spin up headless browsers to inspect JavaScript issues without manual intervention.
- Architectural shifts are expected toward modular or service-based designs with clear API boundaries to manage interdependencies, moving away from large mono-repos, while specific frameworks like Ruby on Rails are favored over languages with less training data like Rust or Elixir, though the latter's success rate may improve soon.
- Model-specific strategies include using Gemini for whole codebase indexing and planning, Sonnet 3.7 for implementing code changes, and testing GPT 4.1 weekly, with an expectation that different models will succeed or fail based on specific contexts and that advice will change weekly as the state of the art evolves.
- Input methods will diversify to include voice tools like Aqua for instructions at 140 words per minute, screenshots for bug demonstration, and downloading API documentation to subdirectories to overcome the patchiness of online web documentation access.
- Continuous evolution of the codebase is planned through frequent refactoring once tests are implemented, relying on LLMs to identify repetitive code, keeping files small and modular, and utilizing tool-specific rule files to enhance model effectiveness.