How To Get The Most Out Of Vibe Coding | Startup School
- Source
- youtube.com
- Author
- Y Combinator
- Date
Context
In a Startup School video, Tom, a partner at Y Combinator, argues that , meaning building software mostly by instructing an AI, improves when you borrow habits from professional engineering. The advice comes from Tom's own experiments on side projects and from comments by founders in the YC batch, so it is experience rather than measured results.
Tom's core habits are to draft a written plan with the AI and work through it section by section, commit working code to version control, and keep high-level tests that click through the app, since AI tools sometimes change unrelated logic. When a fix fails, Tom suggests resetting to the last good commit instead of stacking more prompts on top, because failed attempts leave layers of unneeded code. Tom also recommends writing standing instructions in the tool's rules file.
- YC partner Tom recommends drafting a plan with the in a markdown file, cutting or marking features as won't-do, then implementing one section at a time, running tests, committing to Git, and having the AI mark the section complete.
- Tom finds repeated prompting piles layers of bad code onto a fix, so once a solution works he resets Git and feeds that solution to the AI on a clean codebase. He does not yet trust the tools' built-in revert features.
- Tom prefers high-level, end-to-end integration tests over the low-level unit tests LLMs default to, because they catch the models' habit of changing unrelated logic for no reason.
- Instead of pointing agents at online docs or an server, Tom downloads API documentation into a project subfolder and tells the to read it first. For complex features he builds a standalone reference implementation in a clean codebase.
- Tom attributes strong AI results on Ruby on Rails to its 20 years of conventions and consistent training data, noting friends had less success with Rust or Elixir. At recording time he favored Gemini for planning and Sonnet 3.7 for implementation.
- vibe coding — Building software by describing what you want to an AI in plain language and steering the result, instead of writing every line yourself.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
Checking sign-in…
Loading comments…






