Notes on Mitchell Hashimoto's AI workflow

Mitchell Hashimoto describes how he works with AI in an interview from the Zed team. The conversation stays grounded, occasionally referring to actual commits.

Note he works on Ghostty, a terminal emulator written in Zig. Not necessarily where AI shines the most at this time.

Read the "AI Workflow" sectionAI Workflow

• Tries zero-shots to triage issues. He doesn't necessarily keep the code; it just helps him assess the issue fast (uses a Claude command like /issue 123).
• Mostly he describes a technical solution and lets the AI write the code. He usually has a clear picture of what he wants and asks for just that. Mentions trying "hail mary"s, looking at what the agent produces in a zero-shot prompt.
• Writes a lot of code comments (he always did). This seems to have a good side effect: the AI can spot inconsistencies between the comment and the code.
• Asks the AI: did I miss anything? even on non-AI commits. Thinks about doing something like this in CI.
• He typically refactors by keeping both implementations fully working. Schematically, two folders:

  code-old/
  code/
  

  Asks the AI to check that they do the same thing.

Read the "What AI is good/bad at" sectionWhat AI is good/bad at

• Good at refactoring. It sometimes picks up comments from other files referencing the code and updates them—something a human might have missed.
• Not so good at architecting. Bad at performance: it knows about data structures but not in the context of a project.
• Mentions a piece of SwiftUI view code: 95% AI-generated from a screenshot of Zed's command palette.
• Uses it mostly for what he considers junior/mid-level topics.
• New MacOS features = LLMs are bad.

Read the "AI Usage" sectionAI Usage

Mitchell mentions there is a learning curve to AI tools too:

what tool have you ever adopted, where you were immediately faster [vim helpfully mentioned by the interviewer here]

You sort of have to ramp up in ability before you start to reap the benefits

At some point he mentions AI-generated code in the last two months is "still probably less than 25%", and less than 1% of the entire Ghostty codebase¹.

In the quest for a good dev workflow with AI, this was a helpful video.