Best Agent Skills, Ranked by Verdict

The best agent skills by evidence are Designer Skills Collection (78), Remotion Best Practices (77) and Animate (77). The ranking orders the agent skills we have reviewed by verdict score, never by stars or installs. Agent skills here means SKILL.md packages that AI agents load, not human career skills.

What are the best agent skills?

The best agent skills in our verdicts are Designer Skills Collection (78), Remotion Best Practices (77) and Animate (77). Ten of the skills we have reviewed earn Recommended. None reaches Highly Recommended or High confidence.

• Review date: every skill here was reviewed on 2026-09-30.
• Verdict labels: 10 Recommended, 10 Recommended with caveats, 18 Conditional and 1 Not Recommended.
• Safety grades: 17 A, 12 B and 10 C, with no grade F in this set.
• Confidence levels: 27 Moderate and 12 Low, with no High.
• Top-20 average: 73.55 points across the ranked rows below.

The other skills we have reviewed appear in every reviewed skill, sorted by score.

Skill packs appear below the ranked table without a position. A pack goes here when one of its members holds its own verdict.

Jakub Krehel's pack scores 70. Its member Interface Review scores 68 on its own verdict.

How were these agent skills ranked?

The score gives the worst failure the most weight. Each skill gets seven universal measures: effectiveness, reliability, safety, efficiency, compatibility, maintainability and setup friction. Four to eight measures for its job sit on top. The safety grade comes from reading every file. Stars and installs never move a rank, and payment cannot buy one. The methodology page explains how we score and grade skills. The independence policy explains why no skill can pay for a position.

Popularity leaderboards rank in other ways. The leaderboard at linkly.ai uses a Worth-Installing Score combining popularity, momentum, reputation, maintenance and publisher trust. The skillleaderboard.com site sorts skills by stars and trend. The AWS post on dev.to orders its picks by stars. That post notes that "stars are a popularity signal, not a quality review".

What changed since our review?

GitHub showed these changes on ranked rows by 2026-10-21. Each verdict stays as published until a re-review.

• Brainstorming and Systematic Debugging: obra/superpowers released v7.0.0 on 2026-10-10. We reviewed v6.4.2.
• Using Neon: a commit to the skill path landed on 2026-10-15 (f333dd0).
• Web Design Engineer: commit 6c902cb on 2026-10-10 moved cover images, a low-impact change.
• Claude API: commits 9d63080 and dbd4588 landed on 2026-10-09.
• UI Skills: commits on 2026-10-20 add a create-canvas skill (32afd57, PR #104). We reviewed v0.2.3 at ebf5f26.
• Find Skills: the vercel-labs/skills CLI reached v1.7.4 on 2026-10-19. Its SKILL.md is unchanged since 2026-07-10.
• Claude SEO: v2.4.2 shipped on 2026-10-04 with README edits. We reviewed v2.4.1.
• Jakub Krehel's Interface Skills: commits on 2026-10-12 and 2026-10-13 add technical-draw work. We reviewed 267330e.
• Other ranked rows: no change after review, apart from README edits on Humanizer.

Each re-review is logged under verdict corrections and updates.

Which are the top 10 agent skills, one by one?

The top 10 agent skills all score 76 or higher, and all 10 have a verified Claude Code run. Each card states the verdict, safety grade and confidence. It names the main limitation at review.

1. Designer Skills Collection (Recommended, 78, B, Moderate)

Designer Skills Collection packs research, spec, handoff and critique frameworks with named UX laws. Its maintainer fixed every 2026 load, packaging and routing defect, including #22, #27 and #48. Outside fixes get merged too. The skill is weaker on visual taste than on process. No measured comparison against an unaided agent exists. Grade B reflects seven Python and shell maintenance scripts that users never run. Its rivals sit in the full design skills ranking.

2. Remotion Best Practices (Recommended, 77, B, Moderate)

Remotion Best Practices is the framework vendor's own guidance, maintained almost daily. Reported broken imports (remotion#11669) were fixed within about a day. A side-by-side test found its compositions API-correct but visually bare. A stale global install silently overrides the project copy in Claude Code (remotion#11809, open). The only network endpoint is MapTiler, called with a user-supplied key. Its one GPT run used a Codex-style client, so we do not count it as a Codex run.

3. Animate (Recommended, 77, A, Low)

Animate opens with a gate that asks whether the interface needs motion at all. Fixed easing, duration and spring tables follow, plus a dozen recipes. It refuses motion on high-frequency actions. Issue #28 measured 90% trigger recall on Claude Haiku 4.5. Every miss wrote finished-looking code with invented curves. No recipe includes the reduced-motion block its own Hard Rule 4 requires. Its evidence base is rated Thin, so confidence stays Low.

4. Emil Design Engineering (Recommended, 76, A, Moderate)

Emil Design Engineering reviews UI animation code against concrete easing, duration and transform rules. Tessl measured a 1.20x lift on Claude Sonnet 4.6. The Motion useSpring example never tracks the pointer (issue #26, PR #7). The checklist flags durations over 300ms, while its own table allows 500ms for drawers. About 6,800 tokens load on every use, and the description names no triggers.

5. Frontend Design (Recommended, 76, A, Moderate)

Frontend Design is one 71-line SKILL.md with no scripts or network calls. Its measured file size per load is 2,348 tokens. Tessl's eval on Sonnet 4.6 measured a 1.27x average lift over 3 scenarios. Element-level defaults still slip through: issue #1808 found pill buttons on 10 of 21 pages. The skill gives no guidance for briefs bound to an existing design system. The comparison page answers Frontend Design or UI/UX Pro Max?

6. Brainstorming (Recommended, 76, B, Moderate)

Brainstorming is the design gate from obra's Superpowers collection. The agent clarifies intent, compares approaches and gets sign-off before it writes code. Users credit it with better-scoped features. Runs are verified in both Claude Code and Codex. Approval gates were skipped on Codex Desktop (#2372). The default spec path commits files into docs/ (#1690, #1246).

7. Systematic Debugging (Recommended, 76, B, Moderate)

Systematic Debugging enforces root cause first. Its rules include no fix before Phase 1, one variable at a time and a failing test first. The maintainer fixes real reports within days. At v6.4.2 the multi-layer example prints the IDENTITY secret it checks (#2375, fixed only on dev). The waitFor snippet does not compile (PR #1839, open). Runs are verified in Claude Code and Codex.

8. Using Neon (Recommended, 76, C, Moderate)

Using Neon gives accurate Neon connection, driver and migration guidance. A symptom-to-check table keeps agents on read-only commands. No code ships with it. The workflows still assume Neon account access, a DATABASE_URL and SQL-capable CLI or MCP tooling. Those needs earn grade C. Routing overlaps with the parent neon skill in a small Haiku pilot (issue #76). Read what a C safety grade means before you install it.

9. Semgrep Rule Creator (Recommended, 76, A, Low)

Semgrep Rule Creator from Trail of Bits drafts one well-tested custom Semgrep rule at a time, taint rules first. A contributor's 5-run-per-arm comparison shows it produces validating, passing rules. Every run fetches seven external docs, which adds cost and network dependence. Local references lack advanced taint features such as pattern-propagators. Its measured file size per load is 1,830 tokens. Evidence is rated Thin, which keeps confidence at Low.

10. Terraform Skill (Recommended, 76, A, Low)

Terraform Skill holds Terraform and OpenTofu work to plan-before-apply and plan-before-destroy guards. A targeted destroy wiped Route53 records in issue #42. PR #43 then added a plan-destroy-first rule with explicit confirmation. The rules stay prose, with no hook enforcing them. SKILL.md costs about 5,100 tokens by our estimate. No independent benchmark of its output exists yet.

Are the most popular agent skills the best ones?

No, the most popular agent skills are often not the best ones. UI/UX Pro Max has 131,959 stars and scores 68. Designer Skills Collection scores 78 with about 2,800 stars.

Stars are snapshots taken at review on 2026-09-30. Skill Creator and Frontend Design share the anthropics/skills repo and the same snapshot. Install counts are as listed by linkly.ai on 2026-10-21 and change daily.

Find Skills has the most listed installs, at 3,939,387, and scores 70. Skill Creator sits in a repo with 179,153 stars and scores 57. Using Neon sits in a repo with 97 stars and scores 76. Popularity measures attention. A verdict weighs evidence about what the skill does once loaded.

Stars and installs still carry some signal. A large user base surfaces bugs faster, and several top verdicts cite user reports. That evidence enters our score through the issues themselves, never through the raw count.

Which agent skills should you avoid?

Avoid Landing Page Design (50), the only Not Recommended agent skill in this set. Use Skill Creator (57) with care. You will find OpenAI's two Figma skills (58, 59) in a repo deprecated in June 2026.

• Skip Landing Page Design (Not Recommended, 50). Its content rules ask for organic-looking figures such as 47.2% and believable invented brands. The fix in PR #1 is unmerged.
• Verify Skill Creator's eval results by hand (Conditional, 57), as the Skill Creator verdict explains. The trigger-eval harness returns false negatives (#1721, #1352, #1412). It counts crashes as passes (#1478).
• Replace OpenAI's Figma skill (Conditional, 58) and Figma Implement Design (Conditional, 59). The openai/skills repo was marked deprecated in June 2026. The maintained successor is Figma's plugin in openai/plugins.
• Check Content Research Writer (Conditional, 58, Low confidence) before you trust its examples. They model invented figures and placeholder citations, including a fictional MIT quote. No agent run is verified.

Skill authors get a better route in how to build and test your own skill.

Ten of the skills we have reviewed carry safety grade C. Examples include Using Neon, Claude SEO, Impeccable and shadcn/ui. Grade C means the skill needs credentials, network access or broad permissions to do its job. It does not mean the skill is malicious. Follow how to vet a skill before installing for any grade C skill.

What are the best agent skills for coding and developers?

The best agent skills for coding are Brainstorming and Systematic Debugging (76), both verified in Claude Code and Codex. Terraform Skill, Semgrep Rule Creator and Using Neon (76) cover infrastructure, security and databases.

• Planning and debugging: Brainstorming (76) and Systematic Debugging (76), both from obra's Superpowers.
• Infrastructure and data: Terraform Skill (76), Using Neon (76) and Supabase Postgres Best Practices (66).
• Security work: Semgrep Rule Creator (76), which writes and tests one custom rule at a time.
• APIs and video: Remotion Best Practices (77) and Claude API (72).
• Tooling: Find Skills (70), which suggests skills from the skills.sh catalogue.

Only Brainstorming and Systematic Debugging hold verified Codex runs in this group. Every other coding pick is verified in Claude Code alone. The Claude Code hub ranks coding skills verified in Claude Code.

Supabase Postgres Best Practices scores 66 and sits outside the top 20. It earns its place here as the second database skill. Claude API scores 72 and keeps current model IDs and prices in front of the agent.

Other lists name Matt Pocock's skills, GStack and the Karpathy guidelines. We have not reviewed them, so they carry no score here.

What makes an agent skill good?

A good agent skill has a short entry file, a clear trigger description and a measured result. Our top verdicts share fast maintainer fixes and no hidden network calls. Popularity is not what makes a skill good.

The top 10 verdicts share five patterns.

• Short entry files: Designer Skills Collection loads 1,099 tokens, Semgrep Rule Creator 1,830 and Frontend Design 2,348. These are measured file sizes per load.
• Measured results: Tessl measured 1.27x for Frontend Design and 1.20x for Emil Design Engineering. Issue #28 measured 90% trigger recall for Animate.
• Fast fixes: Remotion fixed broken imports within about a day. Terraform Skill merged a plan-destroy rule (PR #43) after a reported outage (#42).
• Clean file reads: nine of the top 10 hold grade A or B. The top five show no hidden network calls.
• Explicit triggers: a description that names when to load helps the agent pick the skill. Emil Design Engineering names no triggers, a weakness its verdict records.

Remotion's one endpoint is MapTiler, and it runs only with a key the user supplies. That counts as declared, not hidden. Using Neon is the only grade C skill in the top 10.

The guide on what a good SKILL.md contains turns these patterns into a checklist. The tutorial shows how to write and test a skill of your own. New readers can start with what Agent Skills are and how agents load them.

Which agent skills work in Claude, Claude Code, Codex and other agents?

Claude Code has the most verified agent skills: 28 of the skills we have reviewed ran in it. Codex has 7, the Claude app 2 and Cursor 1. Eight skills have no verified agent, only publisher claims.

• Claude hub: the best Claude skills, ranked covers skills with a verified run beyond Claude Code. Humanizer and PPTX ran in the Claude apps.
• Claude Code hub: Claude Code skills ranked by verdict ranks the 28 skills with a verified Claude Code run.
• Codex: seven verified skills, including Brainstorming and Systematic Debugging.

A claim is the publisher's statement, and we cite no run for it. Remotion Best Practices claims Codex, but its GPT run used a Codex-style client. This page does not count that run as a verified Codex run.

The SKILL.md format is an open standard listed at agentskills.io. A skill that follows it loads in any listed client. Its tools and scripts decide whether it works there. Gemini CLI has 8 publisher claims and no verified run in this set. OpenCode has 10 publisher claims and no verified run either.

Which agent skills fit design and writing work?

Two ranked lists group skills by job. Each ranks only the skills whose main job is that work.

The design hub holds 2 Recommended verdicts and 1 Not Recommended. The writing hub holds no Recommended verdict yet. Skills for other jobs are in the full library.

What are the best agent skills on GitHub?

The best agent skills on GitHub often sit in small repos. Designer Skills Collection (78) has about 2,800 stars. The largest repos, obra/superpowers and anthropics/skills, hold verdicts from 57 to 76.

Stars are snapshots taken on 2026-09-30. Repos with several verdicts show the highest snapshot. The last three rows are the smallest repos holding a verdict of 76 or more.

Curated lists, such as those from VoltAgent and heilcheng, collect skills; they do not test them. ComposioHQ/awesome-claude-skills has 76,196 stars, and its one reviewed skill scores 58.

What does the community report about agent skills?

The community reports on agent skills mainly through Hacker News and GitHub issues. Our verdicts cite 17 Hacker News items across 10 skills. Workflow skills draw praise; design skills draw the most criticism.

The 17 Hacker News items split unevenly. Impeccable draws 4. Remotion Best Practices, Frontend Design, Humanizer and Find Skills draw 2 each. Brainstorming, Systematic Debugging, UI Skills, shadcn/ui and Taste Skill draw 1 each. Animate's main community signal is GitHub issue #28 on missed triggers.

Common questions

Is there a full list of agent skills?

Yes. The library lists every reviewed agent skill with its verdict, score and safety grade.

Which agent skills should you install first?

Start with Designer Skills Collection, Remotion Best Practices, Brainstorming, Systematic Debugging and Humanizer. All five have a verified Claude Code run, and no work hub supplies more than two.

Are agent skills safe to install?

Yes, after a check: ten of the skills we have reviewed carry grade C, and none carries F. The vetting guide answers is it safe to install for any single skill.

How are agent skills different from MCP servers?

A skill is an instruction folder the agent loads; an MCP server is a running program exposing tools. The guide on how skills differ from MCP servers compares them in detail.

How often is this ranking updated?

Each verdict carries a review date and the exact commit we reviewed. Every change is logged in the verdict changes log.