How a coding agent talks, measured: blind pairwise picks between Claude Code output styles.
Why it exists
How a coding agent talks is a variable. Ten output styles answered the same questions about the same codebase. The site shows two answers at a time, names hidden, and ten picks give you a ranking and the line that switches the winner on.
What it does
- Blind pairwise picks
- You choose the answer you would rather read, without style names.
- Bradley-Terry ranking
- A ranking with a plain explanation.
- Recorded runs
- Every answer states its harness, version, model and effort.
- Test your own style
- One command adds your style file to the site.
Test your own style
./run.py arm my-style --style-file ~/.claude/output-styles/my-style.md