中文
Same skill, same model. The only difference is whether it gets loaded.

UpliftBench

What an Agent Skill adds, after you subtract the placebo.
Every row comes from three arms — no skill / token-matched neutral text / the real skill. uplift = treatment − placebo. Every row carries N and a 95% CI; the JSON behind this page is index.json.

web-guidance@1.0.0 · grade A · auto activationthe agent decides whether to load it (protocol default)

Skill packageupliftheadroomplacebo95% CIactivationNgrade
addyosmani/agent-skills@301b02b +2.2pp 8.7% -4.4pp [-2.2, 6.7] 17% 30 A

web-guidance@1.1.0 · grade A · auto activationthe agent decides whether to load it (protocol default)

Skill packageupliftheadroomplacebo95% CIactivationNgrade
GoogleChrome/modern-web-guidance-src@3bb177a -2.5pp ⚠️ -9.5% 0pp [-7.9, 2.5] 23% 48 A

web-guidance@1.1.0 · grade A · forced injectionskill body injected into the system prompt; placebo matched by token count

Skill packageupliftheadroomplacebo95% CIactivationNgrade
GoogleChrome/modern-web-guidance-src@3bb177a +5.6pp 15.4% -7.6pp [-2.8, 13.2] 100% 48 A

web-guidance@1.4.0 · grade A · auto activationthe agent decides whether to load it (protocol default)

Skill packageupliftheadroomplacebo95% CIactivationNgrade
multica-ai/andrej-karpathy-skills@2c60614 +4.9pp 8.5% -6.2pp [-3.7, 14.8] 57% 27 B

Tier 0 · facts only (not ranked)

Skill packageprobe resultsgrade
anthropics/skills Tier 0 only · 849,607 tok · largest skill 526,405 tok · 19 skills · defer 84% · monolithic skill · not on the main board (no task set) —
affaan-m/ECC Tier 0 only · 109,921 tok · largest skill 12,192 tok · 47 skills · defer 6% · not on the main board (no task set) —
Nutlope/hallmark Tier 0 only · 409,006 tok · largest skill 409,006 tok · 1 skill · defer 100% · monolithic skill · not on the main board (no task set) —
op7418/Humanizer-zh Tier 0 only · 8,042 tok · largest skill 4,204 tok · 1 skill · defer 0% · not on the main board (no task set) —
KKKKhazix/khazix-skills Tier 0 only · 75,483 tok · largest skill 17,578 tok · 6 skills · defer 83% · not on the main board (no task set) —
mattpocock/skills Tier 0 only · 80,583 tok · largest skill 10,808 tok · 39 skills · defer 28% · not on the main board (no task set) —
obsidianmd/obsidian-skills Tier 0 only · 16,902 tok · largest skill 6,350 tok · 7 skills · defer 0% · not on the main board (no task set) —
cloudflare/security-audit-skill Tier 0 only · 79,967 tok · largest skill 79,967 tok · 1 skill · defer 100% · monolithic skill · not on the main board (no task set) —
obra/superpowers Tier 0 only · 100,774 tok · largest skill 25,738 tok · 15 skills · defer 47% · not on the main board (no task set) —
How to read this table
uplift is treatment − placebo, not minus baseline. headroom is the share of the remaining room that was filled; the placebo column is what length alone is worth.
activation is the share of runs, across the placebo and treatment arms, where the skill actually reached the context. Low activation drags uplift toward zero, because a treatment run that never loads the skill is a baseline run.
Auto activation and forced injection are two different experiments; the same package can have one row each. Those rows cannot be compared with each other, and shipping only the flattering one is how the ecosystem ended up with numbers nobody can reproduce.
Grade A (executable checks) and grade B (a pinned judge) are shown in separate groups, never in one ranking.
⚠️ means the placebo effect is larger than the net gain — most of that number is length.
Generated 2026-11-14T02:55:36.111Z · protocol/1.0.0