Skills registry
The leaderboard is public. So are the evals.
Skills ranked by measured lift — not download counts. Every score is backed by evals you can inspect, case by case.
← Why rank by lift? See how SkillScore worksBuild your own — author, eval, govern ↓AllCodingDataSales & SupportSecurityResearchDevOps✓ Benchmarked only
Top skillsby SkillScore · all categories
lift measured on gemini-3.5-flash · 21+ cases per skill · re-run any skill on your own model with the open runner
| # | Skill | SkillScore ▾ | Lift | Turns Δ | Tokens Δ | Rating | Teams |
|---|---|---|---|---|---|---|---|
| 1st | sql-result-contract ✓ benchmarked data | 91 | +66% | −5% | +2% | ★ 4.9 | — |
| 2nd | secret-scanner ✓ benchmarked security | 89 | +52% | −4% | −2% | ★ 4.8 | — |
| 3rd | commit-conventions ✓ benchmarked coding | 87 | +35% | −12% | −8% | ★ 4.8 | — |
| 04 | error-triage-protocol ✓ benchmarked sales & support | 86 | +28% | −14% | −5% | ★ 4.7 | — |
| 05 | postgres-migration-guard ✓ benchmarked devops | 85 | +40% | −9% | −4% | ★ 4.7 | — |
| 06 | source-citation-guard ✓ benchmarked research | 84 | +44% | −10% | +6% | ★ 4.7 | — |
| 07 | query-plan-reviewer ✓ benchmarked data | 83 | +38% | −8% | −2% | ★ 4.7 | — |
| 08 | pr-review-checklist ✓ benchmarked coding | 82 | +30% | −9% | −4% | ★ 4.6 | — |
| 09 | incident-postmortem ✓ benchmarked devops | 82 | +34% | −13% | +7% | ★ 4.6 | — |
| 10 | refund-policy-writer ✓ benchmarked sales & support | 81 | +30% | −11% | −4% | ★ 4.6 | — |
| 11 | prompt-injection-shield ✓ benchmarked security | 81 | +33% | −2% | −1% | ★ 4.6 | — |
| 12 | json-strict-output ✓ benchmarked coding | 80 | +22% | −6% | −3% | ★ 4.5 | — |
| 13 | search-query-planner ✓ benchmarked research | 80 | +29% | −8% | +4% | ★ 4.5 | — |
| 14 | pii-redactor ✓ benchmarked security | 79 | +31% | 0% | −3% | ★ 4.5 | — |
| 15 | terraform-plan-review ✓ benchmarked devops | 79 | +32% | −7% | −3% | ★ 4.5 | — |
1st
sql-result-contract ✓ benchmarked
91
+66%
2nd
secret-scanner ✓ benchmarked
89
+52%
3rd
commit-conventions ✓ benchmarked
87
+35%
04
error-triage-protocol ✓ benchmarked
86
+28%
05
postgres-migration-guard ✓ benchmarked
85
+40%
06
source-citation-guard ✓ benchmarked
84
+44%
07
query-plan-reviewer ✓ benchmarked
83
+38%
08
pr-review-checklist ✓ benchmarked
82
+30%
09
incident-postmortem ✓ benchmarked
82
+34%
10
refund-policy-writer ✓ benchmarked
81
+30%
11
prompt-injection-shield ✓ benchmarked
81
+33%
12
json-strict-output ✓ benchmarked
80
+22%
13
search-query-planner ✓ benchmarked
80
+29%
14
pii-redactor ✓ benchmarked
79
+31%
15
terraform-plan-review ✓ benchmarked
79
+32%
illustrative sampleSee the live ranking →
Browse the catalog40k+ skills · card view
csv-schema-inference ✓ benchmarked
data
78SkillScore
Infer and pin a column schema from messy CSVs instead of guessing per file.
+27% lift★ 4.4view evals →
ticket-summarizer ✓ benchmarked
sales & support
78SkillScore
Compress a long thread into what the next agent needs to act.
+25% lift★ 4.5view evals →
page-summary-extractor ✓ benchmarked
research
77SkillScore
Pull the claim, the number, and the date out of a page — skip the nav.
+23% lift★ 4.4view evals →
test-case-generator ✓ benchmarked
coding
77SkillScore
Turn a function’s edge cases into runnable tests, not prose about tests.
+24% lift★ 4.4view evals →
dependency-audit ✓ benchmarked
security
76SkillScore
Read an advisory against your lockfile and say whether you’re actually reachable.
+25% lift★ 4.3view evals →
k8s-manifest-lint ✓ benchmarked
devops
76SkillScore
Flag the missing limits, probes, and rollout settings before apply.
+24% lift★ 4.3view evals →
schema-linter ✓ benchmarked
data
77SkillScore
Lint output schemas for missing or drifted required fields.
+26% lift★ 4.4view evals →
support-triage ✓ benchmarked
sales & support
75SkillScore
Route support conversations to the right resolution path.
+21% lift★ 4.6view evals →
web-scraper-toolkit caution
research
73SkillScore
Structured extraction helpers — listed with a caution note.
+22% lift★ 4.2view evals →
changelog-writer ✓ benchmarked
coding
72SkillScore
Draft release changelogs from your merged PRs.
+19% lift★ 4.1view evals →
authz-boundary-check ✓ benchmarked
security
74SkillScore
Check every new route for the tenancy filter it’s supposed to carry.
+23% lift★ 4.3view evals →
runbook-writer ✓ benchmarked
devops
71SkillScore
Turn a fix you just made into steps the next on-call can follow.
+16% lift★ 4view evals →
dbt-model-conventions ✓ benchmarked
data
74SkillScore
Keep model names, tests, and materializations to one house style.
+18% lift★ 4.2view evals →
tone-consistency-guard ✓ benchmarked
sales & support
73SkillScore
Hold replies to one voice across a team of agents and humans.
+17% lift★ 4.2view evals →
rag-chunk-strategy ✓ benchmarked
research
75SkillScore
Chunk on structure instead of character count, then check what retrieval returns.
+20% lift★ 4.3view evals →
illustrative sampleBrowse all skills →
SkillScore
One number, four signals — and one of them can't be bought.
Benchmark lift at 32% (pass rate with the skill minus without), live pass rate from real runs at 32%, an AI-judged quality review at 16%, and adoption at 20% — which grows logarithmically and is capped by org diversity, so volume from a single workspace can’t buy a rank. A missing signal isn’t scored as zero; the remaining weights rescale.
SkillScore · commit-conventions
Benchmark lift · 32%+35%
Live pass rate · 32%86%
AI rating · 16%A
Adoption · 20%—
SkillScore87
no adoption signal yet — the other three weights rescale to 100%
SkillSafety
Nothing lists without clearing the gate.
Every submission runs a three-stage pipeline — static scan, AI security review, content check — and carries its band. Blocked skills never appear.
SkillSafety bands
passedcleared all three stages
cautionlisted with the reason shown
blockednever lists
Skill router
Know which skills your agent actually uses.
smart_route() ranks a shortlist, then compares what you offered against what the agent activated. Under 40% flags menu bloat — and shows which skills to cut.
Offered vs activated · last 7d
Offered12 skills
Activated5 skills
menu bloat42% activation — trim 7
How a skill earns its badge
01Scanned02Security review03Content check04Benchmarked05Band + score published
Build on it
Your team’s skills, not just everyone else’s.
Fork a public skill or write your own, prove it works on your cases, and ship it to your team — on the same rails that rank the public registry. Authoring, evals, versioning, access: skills governance without building the system.
Your eval · refund-policy-writer
Pass rate without54%
Pass rate with84%
Lift+30 pts
workspace-only · illustrative — run it with the open eval runner
From your draft to your team’s router
01Author or forkStart from a benchmarked skill or a blank one. The skill format, linter, and local safety scan are built in.
02Eval before you shipRun the same harness that scores public skills against your own cases — pass rate before and after, and the exact lift.
03Share with your teamPublish to your workspace. Your agents pick it up via smart_route() — and real usage feeds back into the score.
04Governed by defaultThe same safety gate and pinned versions as the public registry. Governance runs on the rails — not as a process you police.
Skills governance
Everything you’d build internally — already here.
Keeping a team’s skills authored, evaluated, safe, versioned, and shared is real work — skills governance. In-house it’s a stack you build and maintain. Here, it’s what the registry already does.
Skills governanceBuild it in-houseOn DecimalAI
Eval harness — with / without runsyou build it✓ the same harness that ranks the registry
Safety review on every versionyou build it✓ SkillSafety gate included
Quality ranking nobody gamesyou build it✓ SkillScore — measured lift
Versioning & pinningyou build it✓ pinned versions built in
Sharing & access controlyou build it✓ workspace publishing
the upkeepweeks of engineering · yours to maintainpip install decimalai · today
Publish your first skill →
FORK & EVAL ON FREE · TEAM PUBLISHING FROM CORE $49
Adopt the skills that are proven to help.
Browse free — no sign-up. Two lines to instrument your agent.
Browse the registry