Four Modes of AI-Assisted Coding (Don’t Just Vibe to Production)

Source: This is how you should use AI to code | don’t just vibe to production. - YouTube

Key Ideas

  • Not all code should be treated equally. Distinguish between code that matters (production) and throwaway code (prototypes).
  • Inspired by Andrej Karpathy’s principle: use AI differently depending on code criticality.

The Four Main Modes

  1. Throwaway / Vibe Coding Mode

    • Prototyping only, code is not meant to last.
    • Large, exploratory prompts. Useful for UI flows or quick proof-of-concepts.
    • Use tools like Boomerang or Micromanager.

  2. New Project Mode

    • Foundational setup for a new codebase.
    • High emphasis on clean, up-to-date boilerplate.
    • Validate everything runs smoothly—no deprecation warnings.

  3. New Feature Mode

    • Code matters, but can be revised or discarded during iteration.
    • Discuss implementation options in chat before generating code.
    • Good use case for orchestrator modes.

  4. Feature Improvement / Bug Fix Mode

    • Surgical, focused updates to existing production code.
    • Small context, minimal files.
    • High quality required. Frequent commits and context resets.

Pair Programmer Mode

  • Custom prompt config using Gemini 2.5 Pro:
    • Always starts by offering multiple implementation options.
    • Clarifies goals before writing any code.
    • Useful for high-trust, production-level edits.

Core Principles

  • Context Strategy: Include everything relevant but avoid overwhelming with too much context.
  • Incremental Changes: Make distinct, surgical modifications rather than large sweeps.
  • Options Discussion: Always discuss approaches with AI before implementation, don’t accept first solution.
  • Know Your Mode: Consciously identify which mode you’re in before starting.
  • Human Guidance: Build prompts yourself, especially for production code - don’t rely entirely on AI.