Reaction
Do AI skills actually improve the work?
Writing about 16 open-source product-marketing skills, the author says install counts reveal little about whether they work.
TLDR
The author proposes testing the same model on the same job and evidence, with and without a skill. They describe an AI making a literal battle card image instead of a usable sales document. Citing Vercel’s Next.js eval, they say merely making a skill available left the agent’s pass rate at 53%; telling it when to use the skill raised it to 79%.
Combined views
—
1 Source, first seen ago
