How To Get The Most Out Of Vibe Coding | Startup School

The video reframes "vibe coding" as a measurable engineering skill focused on results, not semantic debates, and offers a practical playbook for working effectively with AI tools. Key tactics include running multiple AI IDEs in parallel to compare outputs, treating the LLM like a professional developer by starting with a written plan in a markdown file, and implementing features section by section with Git commits and resets to maintain a clean codebase. The speaker emphasizes using raw error messages, handcrafted tests, and high-level regression tests as guardrails, while codifying project rules in tool-specific files and downloading API docs for the model to read. For complex features, prototyping standalone before integration, keeping code modular, and choosing tech stacks with abundant training data like Ruby on Rails dramatically improve results. Finally, the advice is to refactor early, lean on tests, and continually experiment with new models as the field evolves rapidly.

• ▶ 0:09 Vibe coding is a measurable skill: experiment with AI tools, apply software engineering best practices, and focus on results rather than semantic debates.
• ▶ 1:05 Workaround stuck AI IDEs by pasting code into the LLM chat UI, and run multiple IDEs like Cursor and Windsurf in parallel, comparing outputs to pick the best iteration.
• ▶ 2:06 Treat AI as a new programming language: give detailed context, plan scope and architecture first, handcraft test cases as guardrails, and step back if the model starts rabbit-holing.
• ▶ 4:00 Treat the LLM like a professional developer: start with a written plan in a markdown file, not code, and keep refining it as a living document.
• ▶ 5:48 Implement the plan section by section, checking each piece works and committing before moving on, rather than trying to one-shot the whole product.
• ▶ 6:26 Use Git religiously and reset to a clean state whenever the AI goes off track; also have the LLM write high-level end-to-end tests to catch regressions early.
• ▶ 8:58 Paste raw error messages directly into the LLM; for complex bugs, ask it to brainstorm possible causes first, reset between failed attempts, and fix on a clean codebase to avoid piling on junk code.
• ▶ 10:44 Codify persistent project rules in tool-specific instruction files (e.g., Cursor Rules, CLAUDE.md), and download API docs locally, telling the LLM to read them before implementing for much more accurate results.
• ▶ 12:04 For complex features, prototype standalone in a clean project to get a reference implementation, then point the LLM at that reference while reimplementing the feature inside the larger codebase.
• ▶ 12:35 Keep code small and modular—embedding code in a larger codebase works well, and a consistent external API lets you change internals safely as long as tests pass.
• ▶ 13:20 Choose a tech stack with abundant, consistent training data (e.g., Ruby on Rails) because AI performs much better on frameworks with well-established conventions.
• ▶ 15:09 Refactor early and often once tests exist, and keep experimenting with new models week to week since the state of the art changes rapidly.

• ▶ 0:09 Opening & Founder Tips (0:09 - 3:54) - Tom introduces the vibe coding experiments and shares practical founder tips for working with LLM coding tools.
• ▶ 3:54 Professional Developer Workflow (3:54 - 8:58) - Choose the right tool, plan with the LLM, use version control, and write high-level tests for a reliable process.
• ▶ 8:58 Debugging, Docs, and Complex Features (8:58 - 12:39) - Paste errors back into the LLM, codify instructions and docs, use the LLM as a teacher, and prototype complex features standalone.
• ▶ 12:39 Architecture, Iteration, and Outro (12:39 - 16:41) - Keep code modular, pick the right stack, use screenshots/voice, refactor often, experiment with models, and wrap up.