Controlled experiments for agent UI skills
Which UI skill actually makes your agent build better pages?
Every page here was generated by the same model from the same product brief. The only thing that changes from row to row is the skill in the system prompt, and the first row uses no skill at all as the control.
- Skills compared
- Product briefs
- Pages generated
- Model
The matrix
One row per skill, one column per brief. Start with the control row, then compare any skill against it. Select a skill to see all runs.
How the comparison works
SkillArena isolates one variable: the skill. Everything else stays fixed, so the differences you see come from the skill and nothing else.
Same brief, same model
Each skill receives the exact same product brief and runs on the same model (the same model). Only the skill text in the system prompt changes.
Three runs per pair
Every skill and brief pair is generated three times, so one lucky or unlucky run does not decide the result. The thumbnail shows one of them.
Full results on the site
Skill names and thumbnails link to the complete results on skillarena.dev, with every run and the live pages.
Join the waitlist
Get an email when new skills, briefs or models are added to the arena.
- Only updates about new results, no other mail
- One field and a role, nothing else
- Unsubscribe from any email in one click