Question 1

How does SkillzBench benchmark a SKILL.md skill fairly?

Accepted Answer

Every skill is tested in a sandbox with the same model, run twice — once with the skill and once without — and the Lift is published in percentage points. Because the model and conditions are identical, the Lift reflects the skill's actual contribution, not prompt variance or model differences. Lift is always shown separately from trust ratings so performance and trust never get conflated.

Question 2

What does the 0-100 trust score actually measure?

Accepted Answer

The trust score blends five signal types: security, quality, license, maintenance, and community. It sits on top of the Verify pipeline, which pins every version to a commit SHA and runs validation, license-check, security-scan, and quality-judge steps. Money never touches scores — ratings can't be bought.

Question 3

Does SkillzBench handle script-bearing skills, or only instruction-only SKILL.md packages?

Accepted Answer

Launch scope is instruction-only skills — SKILL.md plus text references. Script-bearing skills are listed link-out only, so you can discover them but they aren't benchmarked or equipped through the platform. This keeps the verified surface area auditable and safe.

Question 4

When can I actually use SkillzBench, and what will it cost?

Accepted Answer

SkillzBench is pre-launch and actively being built — the waitlist and newsletter are live today. Planned pricing includes a generous free tier, Pro at roughly $9-15/month for more agents and unlimited enabled skills, Teams at roughly $15-25 per seat/month for org loadouts, allowlists, and audit, and a Creator tier around $5-10/month for verified authors, plus benchmark-as-a-service for priority runs.

Question 5

How is this different from just browsing skills on GitHub or a community skill registry?

Accepted Answer

GitHub shows you a README; SkillzBench shows you evidence. Every version is pinned to a commit SHA, security-scanned, license-checked, and quality-judged, then benchmarked with-and-without in a sandbox with Lift in percentage points. One MCP server with per-agent loadouts means your agent only sees what you chose, and upstream changes trigger re-scan, re-benchmark, and notification — none of which a raw GitHub search provides.

Verified Agent Skills for Freelancers