Vibe Coding at Scale: Customizing AI Assistants for Enterprise Environments - Harald Kirshner,

Introduction to Vibe Coding 00:46

  • VS Code Insiders is a pre-release version built daily, offering the latest features, while the main VS Code releases monthly.
  • "Vibe coding," as coined by Andre Kaparthy, is an approach that embraces exponential code generation and shifts focus away from the code itself.
  • It's described as starting in a "fun chaos mode" but aims to lead to "professional zen vibes."

Levels of Vibe Coding 01:16

  • YOLO Vibe Coding is not meant for production, focusing on speed, instant gratification, creativity, and fast learning.
  • Structured Vibes introduce guard rails, leading to maintainability and more sustainable code, suitable for enterprise environments.
  • Spec-driven Vibes scale the concept to large codebases, emphasizing reliability and velocity while ensuring generated code fits the existing codebase.

YOLO Vibe Coding in Practice 03:04

  • The approach is outcome-focused, iterating through language by talking to the AI and auto-accepting its suggestions.
  • If something goes wrong, the process involves resetting or undoing completely and continuing.
  • YOLO coding is suitable for rapid prototyping, proof of concept, exploring new ideas, fostering creativity, learning new tech stacks (e.g., Rust), and personal projects.

VS Code Features for Vibe Coding 04:20

  • VS Code Insiders is recommended for access to the latest features, including an "agent mode" (not yet default).
  • An "auto-approve" chat option can be enabled, even for specific workspaces, to bypass confirmation buttons during rapid iteration.
  • VS Code includes a built-in microphone button for local model interaction, ensuring no data leaves the machine.
  • The "create new workspace" tool facilitates easier vibe coding by setting up environments.
  • The demo used Cloudin 4, noted for its good UI generation, contrasting with other models that produce wireframe-like or blockchain-landing-page designs.
  • The YOLO toolbox includes agent mode, new workspace flow, voice dictation, visual context, auto-accept (including "after delay"), auto-save for live updates, and undo/revert options.
  • The AI can analyze terminal output, automatically fix errors, and install missing dependencies.

Structured Vibe Coding for Enterprise 10:03

  • This level involves using templates with a consistent tech stack, which can be internal, containerized, or accessible to non-technical users.
  • Templates come with instructions to guide Copilot to adhere to specific conventions already in use within an organization.
  • Custom tools (like MCP) can be integrated to provide the AI with necessary information and context, leading to more reliable and consistent output than YOLO mode.
  • This approach is ideal for bootstrapping screen-fill projects and better handles less popular or off-mainstream technologies by providing explicit instructions.

Advanced Customization and Best Practices 12:35

  • VS Code allows defining "chat modes" (e.g., a test-driven development agent) with specific tools and system prompts that guide the AI's behavior.
  • Reusable "prompts" can be created for teams to standardize common tasks, like generating new landing pages.
  • "Instructions" (githubinstructions) can outline specific parts of the codebase to provide the AI with context.
  • VS Code's access to problems and tasks allows the AI to fix code mistakes as they occur, providing powerful context.
  • The "spark icon" in the commit view highlights the importance of committing workable code frequently to avoid losing progress or introducing regressions.
  • Key practices include reviewing, starting with errors, committing often, and pausing the agent to inspect generated code.

Spec-Driven Vibe Coding 14:55

  • This highest level of vibe coding involves starting with a clear specification, creating a plan, and then implementing that plan.
  • It leverages prompts, more MCPs (custom tools), and task-specific tools to scale to large projects.