Author deletes all agent skills after benchmarks show poor results

After reviewing benchmarks showing that public agent skills frequently fail to improve task success while driving up token costs and introducing security vulnerabilities, the author decided to delete all their agent skills. Citing findings from SkillsBench, SWE-Skills-Bench, and observations from figures like Anthropic and Andrej Karpathy, the author notes that while curated skills can help, many public skills merely repeat documentation or overconstrain the model. Open article →