Vibing a Non-Trivial Ghostty Feature
TLDR
- Mitchell Hashimoto shipped Ghostty’s macOS unobtrusive update notifications using AI-assisted agentic coding across multiple scoped sessions.
Key Takeaways
- The feature was triggered by a Ghostty update prompt interrupting a live OpenAI keynote demo; the fix was non-modal update UI in the titlebar or window overlay.
- Hashimoto used Amp with an “oracle” read-only subagent for planning, always generating a written plan before letting the agent write code.
- When the agent hit a persistent layout bug across four targeted sessions, he pivoted rather than continuing to prompt: the update notice moved from titlebar to bottom-right window overlay.
- AI was used for “zero to one” UI prototyping, fill-in-the-blank scaffolding, and documentation; human coding followed each session to clean, reorganize, and validate.
- Code the author couldn’t understand after AI produced it was backed out; documentation sessions were treated as mandatory anti-slop checkpoints.
Why It Matters
- The post is a rare full-transparency account: every agentic session shown unedited, including failures, pivots, and token costs.
- The workflow pattern is concrete and repeatable: plan with oracle, prototype in small chunks, clean up before moving on, never ship code you can’t explain.
- Hashimoto’s framing that AI is a liability during persistent bugs, not a solution, is a direct counterpoint to replace-the-engineer narratives.
Mitchell Hashimoto · 2025-10-11 · Read the original
