Three popular agent test skills, including one with 250k stars, cost 26-41% more and did not improve correctness
Dan Luu·high signal
The same study evaluated four skills: the ECC Rust test skill (250k GitHub stars, 38k forks), the official Hegel skill, the Trail of Bits property-test skill, and a custom skill the author wrote. The Hegel skill's SKILL.md plus its linked Rust reference exceed 20k tokens and was read in 157 of 160 runs, adding 26-41% cost with no correctness gain. The author's own skill scored highest, and he attributes the difference to writing instructions that nudge the agent away from its default behavior rather than reading like a tutorial.