What Is a Good Lift Score for a Skill?

The Lift Problem: Why Skill Scores Mislead

You install a skill, and your agent gets worse.

This is not a hypothetical scenario. It is a documented, recurring failure mode in the AI agent ecosystem.

Thousands of agent skills are published weekly, yet most lack independent verification.

Polished SKILL.md files often look impressive but actively degrade performance.

The core problem is that skill quality is not visible from the file itself.

You cannot tell if a skill helps or hurts without testing it.

This is where the concept of lift becomes critical.

Lift measures the actual performance gain an agent gets from a specific skill.

It is the difference between running with the skill and running without it.

Without this metric, you are guessing about your agent's reliability.

Research shows that negative lift is common and real.

In one major benchmark, 13 out of 87 tasks regressed when skills were added.

Self-generated skills performed 8.1 to 11.5 points below the no-skill baseline.

This means adding a skill is not automatically an improvement.

It is an empirical question that requires testing.

The "good" lift score is not a single universal number.

It depends on the model, the harness, and the specific task domain.

Lift varies dramatically across different configurations.

You might see a +4.1 point gain in one setup.

Another setup might show a +25.7 point gain.

There is no fixed threshold that works for every agent stack.

Instead, you need a statistically meaningful positive gain validated on your own system.

This is why independent verification matters more than ever.

The skill ecosystem is growing explosively, and so is the risk.

Security researchers found that 13.4% of skills contain critical security issues.

A skill can be useful and dangerous at the same time.

You need to measure both performance and safety.

SkillzBench is building a free directory and MCP tool to solve this.

It benchmarks every skill in a sandbox, with and without the skill.

This approach separates helpful skills from harmful ones.

It also verifies security and quality independently of performance scores.

You can browse and test free skills on SkillzBench to see the lift for yourself.

The goal is simple: only keep what measurably helps your agent.

Discard the rest.

What the Benchmark Data Actually Says: Lift Is Contextual

If you came here looking for a magic number, the research has an inconvenient answer: there isn't one. The most rigorous benchmark work on agent skills concludes that skill efficacy is "an empirical property of a specific agent stack rather than a universal constant" — meaning lift depends entirely on the model, harness, and task domain you're actually running (arXiv preprint).

That said, the data does give useful anchors. Across 87 tasks and 18 configurations, curated skills raised the average pass rate from 33.9% to 50.5% — a +16.6 percentage point gain, with a normalized gain of 25.5% (SkillsBench research). But the average hides enormous spread: configuration-level lifts ranged from just +4.1 pp to +25.7 pp.

Domain matters just as much as model. An earlier version of the benchmark measured domain-specific improvements running from +4.5 pp in software engineering to +51.9 pp in healthcare (earlier SkillsBench results). A skill that looks mediocre in one domain can be transformative in another.

Why stronger models often benefit less comes down to math. SkillsBench uses a normalized gain metric — g = (pass_skill − pass_vanilla) / (1 − pass_vanilla) — which accounts for how much headroom a model has (methodology details). The pattern is consistent across the leaderboard:

• Haiku 4.5 gained +21.3 points from a weak 8.8% baseline, while the stronger Opus 4.8 gained only +8.4 points from a 45.7% baseline (SkillsBench).
• Mid-tier models produced the largest lifts — GLM 5.1 on OpenHands gained +25.7 pp — while strong baselines like Gemini 3.5 Flash gained only +7.1 pp (benchmark data).
• The paper's own conclusion: "high base capability does not imply high Skill leverage" (arXiv).

This is why context beats a single threshold. A +5 pp lift on an already-strong model may represent excellent leverage; the same number on a weak baseline suggests the skill is underperforming. The practical standard, per practitioners in the field, is to accept a skill "only when a matched with-skill versus without-skill evaluation shows meaningful lift" (Arize's guidance).

That paired, same-stack evaluation is exactly the methodology SkillzBench is building into its benchmarking pipeline: testing every skill in a sandbox with the same model, with and without the skill, and publishing the resulting Lift in percentage points. Because lift is contextual, the number only means something when you know the conditions it was measured under.

When Skills Hurt: Negative Lift and the Paired-Evaluation Standard

Not every skill makes an agent better — some make it measurably worse. That uncomfortable finding sits at the heart of the lift question, and it's why "good" can never be a single number.

The SkillsBench research is blunt about this: benchmark data shows that 13 of 87 tasks actually regressed when skills were added, with an earlier benchmark version reporting regressions in 16 of 84 tasks. In other words, roughly one in six or seven tasks got worse, not better. A polished-looking SKILL.md can still make an agent less reliable, regardless of how convincing the documentation reads.

Self-generated skills are even more troubling. When models wrote their own skills, performance fell 8.1 to 11.5 points below the no-skill baseline across Claude Code, Codex, and Gemini CLI, according to the same research synthesis. Curated, human-authored skills on identical configurations delivered gains of +18.2 to +24.8 points. The conclusion from the SkillsBench paper: models cannot reliably author the procedural knowledge they benefit from consuming.

This is why the practical standard for "good" is methodological, not numeric. As the SkillsBench paper argues, skill efficacy is an empirical property of a specific agent stack — a particular model, harness, and task — rather than a universal constant. A skill that lifts one configuration by 25 points might do nothing, or damage, on another.

The accepted standard, then, is a matched paired evaluation: run the same task, in the same container, with and without the skill, and accept the skill only when the with-skill run shows a statistically meaningful positive delta. Arize frames it as a shipping gate — "accept each change only when a matched with-skill versus without-skill evaluation shows meaningful lift" — because aggregate averages hide per-task regressions.

For teams evaluating skills, that translates into a short discipline:

• Never assume a skill helps — test it against a without-skill baseline on your own stack.
• Watch for negative lift on individual tasks, even when the average looks positive.
• Treat self-generated or auto-summarized skills with particular skepticism.
• Remember that lift and safety are separate axes — a helpful skill can still carry a critical vulnerability.

That last point matters more than teams often realize. Snyk's audit of the ClawHub registry found 13.4% of skills contain at least one critical security issue, so a positive lift score alone tells you nothing about whether a skill is safe to deploy.

This paired, with-and-without testing methodology is exactly what SkillzBench is being built around: benchmarking every skill in a sandbox against the same model, with and without the skill, and publishing the resulting Lift in percentage points — always shown separately from trust ratings, so a helpful-but-risky skill can't hide. The core principle is simple: keep what measurably helps, and discard what does not.

How to Judge a Skill Before You Install It

Before you install a new agent skill, verify it actually improves your workflow. A polished instruction file is not a guarantee of performance; it is a hypothesis that requires empirical testing.

The most reliable way to judge a skill is to prefer focused modules over comprehensive bundles. Research shows that skills with two to three modules deliver a +21.5 point lift, while exhaustive documentation yields only +0.7 points. This suggests that concise, targeted instructions help agents better than dense, all-encompassing guides.

Evaluate skills specifically on your own model and harness stack. The same model can perform differently depending on the interface, with scores shifting by up to 8 percentage points based on the harness choice. A skill that works well in one environment may underperform in another, so generic benchmarks are less useful than targeted tests.

Treat lift and security as separate axes of evaluation. A skill can show positive performance gains while carrying critical vulnerabilities. In fact, 13.4% of skills contain at least one critical security issue, meaning performance does not equate to safety. You must verify both dimensions independently.

Use this checklist to screen skills before deployment:

• Check the module count: Prefer skills with 2-3 focused modules over large bundles.
• Test on your stack: Run the skill with your specific model and harness to see real-world results.
• Scan for security flaws: Verify the skill has been audited for critical vulnerabilities, separate from its performance data.
• Compare with baseline: Ensure the skill provides a measurable lift over the no-skill baseline for your specific tasks.

SkillzBench is building a free directory and MCP tool to help you navigate this process by benchmarking skills in sandboxed environments. This ensures you see the actual lift and security status before you equip your agent, keeping your workflow efficient and secure.

Measuring Lift for Yourself: From Benchmarks to Your Stack

Knowing a skill's lift should not require running your own benchmark lab, yet that is exactly what rigorous measurement demands. The SkillsBench methodology calls for paired evaluation — testing the same task with and without the skill, on the same model, in the same container — because lift is "an empirical property of a specific agent stack rather than a universal constant," not a number you can read off a README.

That is the gap SkillzBench is being built to close. The platform's benchmarking pillar tests every skill in a sandbox with the same model, with and without the skill installed, and publishes the resulting Lift in percentage points. Given that curated skills raise average pass rates from 33.9% to 50.5% (+16.6 pp) across 18 configurations, but individual configuration gains range from +4.1 to +25.7 pp, per-skill paired numbers are the only way to know where a given skill actually lands.

Crucially, Lift is published separately from trust. A skill can measurably help while carrying critical vulnerabilities — research found 13.4% of skills contain at least one CRITICAL security issue — so a single blended score would hide real risk. SkillzBench keeps the two axes apart and reports both.

The verification pipeline behind each listing includes:

• Version pinning to a commit SHA, so what you benchmarked is what you install
• Security scanning and license checks on every version
• Quality judging and validation of the skill package itself
• Re-scanning and re-benchmarking as upstream changes land, with notifications

This matters because the skill ecosystem is moving fast and unevenly. Published counts of new skills jumped from under 50 per day to over 500 in a matter of weeks, and a security audit of one registry uncovered roughly 12% of listed skills were malicious. Speed without verification is how a polished-looking SKILL.md makes your agent less reliable — 13 of 87 benchmark tasks regressed when skills were added.

Once a skill clears both bars — meaningful lift and a clean trust score — equipping it should be simple. SkillzBench's plan is a free directory where you can browse skills, read their measured Lift and trust score, and connect the ones you choose to your agent through a free MCP connector, with per-agent loadouts so an agent only sees the skills and tools you toggle on. Everything stays free; advertising in blog content funds the platform, and ad money never touches scores or listings.

The standard the research sets — keep what measurably helps, discard what does not — only works if someone has already done the measuring. Sandboxed, same-model, with-and-without benchmarking is how "does this skill help?" stops being a guess.

ctaText: Browse skills with published Lift scores — free on SkillzBench socialProofText: SkillzBench benchmarks every skill with and without it on the same model, so you see the real gain before you install.