- Author
- David Bau
- Published
- 16 December 2025
- Kind
- Personal blog post with one worked project and reader comments
- Original
- davidbau.com/archives/2025/12/16/vibe_coding.html
- Subject
- Harnesses and agent loops
- Themes
- Art through code, Handing over control
What it is
David Bau keeps a personal weblog. The post reports an experiment in letting Claude Code rewrite an old fractal viewer page. It draws two rules for working with coding agents and reflects on losing human understanding.
What it says
- Bau splits vibe coding into delegating small tasks and ceding control.
- Bau's page grew from 780 human-written lines to over 13,600 lines.
- Bau reports the agent's version renders a deep zoom in one minute.
- Bau's first rule is to automate tests so the agent checks its own work.
- Bau's second rule is to test the tests, because agents can pass weak tests.
- Tests catch bugs but not bloat, so Bau made one human pass to simplify.
- Bau warns that more complexity can cost humans their understanding of what they want.
What evidence it gives
The evidence is one hobby project, a fractal viewer page. Bau gives line counts and a before and after render. It links to code repositories and git history. There is no controlled comparison or outside measurement.
Interests and limits
This is one project by one author using one coding agent. Bau says the page was tested on a Mac. One commenter saw only a black screen, and Bau replied that it needs WebGPU. The last section is reflective opinion on AI and human understanding.