Vibecoding reports
Vibecoding refers to generating working software (mostly) without reviewing or manually modifying generating code. I’ve listed some writeups of successful application of this approach.
At this point the scale of this approach is likely to be limited to smaller projects but people do report success with projects that span thousands of lines of code.
Mitchell Hashimoto (founder of HashiCorp, author of GhostTTY) wrote a detailed walkthrough of adding a feature to a Swift project with the code mostly generated by AI. There’s a good workflow description and tips, and it looks like he completed the feature significantly faster with AI assistance.
Diwank Tomer has written up Field Notes From Shipping Real Code With Claude. He describes three modes of vibecoding:
- Playground mode - for one-off scripts, hacks and quick prototypes
- Pair programming mode - for projects under 5k lines of code
- Production mode - working on large existing codebases
The last mode is less likely to be successful but that’s why it’s particularly interesting to read about Diwank’s approach.
Maciej Trebacz estimates that 95% of his game jam entry Tower of Time was generated with Augment Code and Cursor. See also PROMPTS.md for details of dev process and prompts used.
Indragie Karunaratne reports that Context, their native macOS app for debugging MCP servers, was almost 100% built by Claude Code. It required ~20k LOC.
In a very niche application of AI dev tools, Dmitry Brant used Claude Code to modernise a 25-year old Linux kernel driver. These kinds of esoteric (but useful) project seem to be an area with good potential for AI dev tools being helpful.
A ThoughtWorks experiment with Claude Code describes success in adding support for Python code ingestion in their CodeConcise tool, followed by failure to implement support for JavaScript.
Dreams of Code describes a 30-day vibecoding experiment [video] which resulted in successfully reimplementing a course web app in Next.js and adding features to a Go web app.