← @bgSponge

Noticing myself adopting a certain rhythm in AI-assisted coding (i.e. code I actually and…

Summary

Andrej Karpathy proposes a disciplined seven-step rhythm for professional AI-assisted coding to avoid the low-quality results of "vibe coding."

  1. Context saturation is the first step.

    Users must include all relevant files, such as TypeScript or CSS, into the prompt context, which can be time-consuming in large projects.

    Stuff everything relevant into context

  2. High-level approaches precede code generation.

    Karpathy advises describing a single concrete change and asking for pros and cons of different approaches rather than requesting code immediately.

    Don't ask for code, ask for a few high-level approaches

  3. Manual review enables inline learning.

    The developer should manually pull up API documentation for unfamiliar functions to verify the LLM's output and clarify explanations.

    pull up all the API docs in a side browser

  4. LLMs act as unprincipled interns.

    Karpathy characterizes the AI as a savant with encyclopedic knowledge who nonetheless bullshits, lacks taste, and requires a tight leash.

  5. Professional coding requires defensive pacing.

    Effective AI assistance relies on being slow, careful, and paranoid rather than delegating the entire cognitive load to the model.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →