Skills Night: 62,000가지 방법으로 에이전트가 더 똑똑해지고 있다
Source: Vercel Blog
번역하려는 본문을 제공해 주시면 한국어로 번역해 드리겠습니다.
What We Learned
Tuesday night we hosted Skills Night in San Francisco, an event for developers building on and around the open‑skills ecosystem we’ve been growing since the idea started as a single weekend of writing.
What began as Shu Ding sitting down to document everything he knows about React has grown into 62 000 skills, 2 million skill‑CLI installs, and a community moving incredibly fast.
skills.sh
The Origin Story
“Shu Ding is one of the most talented web engineers I’ve ever worked with. He knows things about React and the browser that most people will never discover.” – [Name]
Last year Shu spent a weekend writing a kind of web bible—a collection of markdown documents that captured everything he knew about React.
We wanted to ship it, but the usual options didn’t fit:
| Option | Why it didn’t work |
|---|---|
| Blog post / static docs | Too slow to surface in next‑gen models (Claude Sonnet 8, GPT‑9) |
| MCP server | Overkill for a simple markdown collection |
Skills turned out to be the quickest way to deliver on‑demand knowledge.
While writing the instructions for installing React best practices, I kept copying the same installation steps for Cursor, Claude Code, Codex, and ten other coding agents—each with slightly different install directories.
So I built a CLI that installs a skill into every major coding agent at once. That CLI became the core of the ecosystem.
We added telemetry to surface new skills as they got installed; this data now powers the leaderboard at skills.sh.
“It’s a package manager for agent context.” – Malte Ubl, Vercel CTO
The whole thing went from idea to production on Vercel in days.
npx skills skills.shNow we track 62 000 skills and make them easy to discover and install with a single command, e.g.:
npx skill installGrowth = Attack Surface
Rapid growth brought quality variance.
Example: A skill that looked clean in markdown contained a Python file that opened a remote shell on install. You’d never catch that without inspecting every file in the directory.
Our response: Security partnerships with GenSocket, Snyk, and [Partner 3] to audit all existing skills and every new one that arrives.
We’re building an Audits leaderboard that provides per‑skill assessments and recommendations. The goal isn’t to lock things down; it’s to let you move fast with confidence. We’re always looking for new security partners who can bring unique perspectives and more trust signals to the ecosystem.
Partner Demos & Key Takeaways
| Partner | Demo Highlights | Core Insight |
|---|---|---|
| Ben Davis (Skills) | Ran a controlled experiment to close the training‑cutoff gap. Tested Svelte remote functions in four ways: 1️⃣ No context 2️⃣ Skills file with docs 3️⃣ Skill pointing to MCP 4️⃣ Code example in project | Every approach with context worked. The no‑context run (forced through a stripped‑down model) produced completely wrong output. Context matters more than medium; skills are the fastest way to provide that context. |
| Evan Bacon (Expo) | Showed native iOS feature upgrades driven entirely by Claude Code using Expo skills. New SwiftUI components, gesture‑driven transitions, and tab‑bar updates were applied automatically. Also demonstrated a work‑in‑progress skill that uses LLDB to read the iOS view hierarchy and auto‑fix keyboard‑handling bugs. | Agents can now drive the whole stack. Expo Go now auto‑fixes every crash as it occurs—a huge win for anyone who’s wrestled with Xcode. |
| Nick Khami (Mintlify) | Mintlify auto‑generates a s |
Source: …
kill for every documentation site they host (Claude Code docs, Coinbase, Perplexity, Lovable, etc.). Traffic to these sites is now 50 % coding agents, up from 10 % a year ago. | Skills are becoming infrastructure. The skill is no longer a manual doc‑team output; it’s a by‑product of well‑structured documentation. | | David Cramer (Sentry) | Built Warden, a harness that runs skills as linters on pull requests via GitHub Actions, treating agents as a static‑analysis layer. | Demonstrates how skills can be embedded directly into CI pipelines. |
“Agents make mistakes. They sometimes tell you you’re absolutely right and then do the wrong thing.” – Guillermo Rauch, Vercel CEO
Shipping quality in the AI era means raising the bar on what those tokens actually produce, not just counting them. Skills give us a way to influence agents, keep them up‑to‑date with framework changes, and make them more token‑efficient by providing a straight path to the right answer.
번역
그들이 호스팅하는 모든 문서 사이트(Claude Code docs, Coinbase, Perplexity, Lovable 등)에서 트래픽이 이제 50 % 코딩 에이전트에 달했으며, 이는 1년 전 **10 %**에서 증가한 수치입니다. | 스킬이 인프라가 되고 있습니다. 스킬은 더 이상 수동적인 문서 팀의 산출물이 아니라, 잘 구조화된 문서의 부수적인 결과물입니다. | | David Cramer (Sentry) | Warden을 구축했으며, 이는 GitHub Actions를 통해 풀 리퀘스트에 스킬을 린터로 실행하는 하네스로, 에이전트를 정적 분석 레이어로 취급합니다. | 스킬을 CI 파이프라인에 직접 삽입할 수 있는 방법을 보여줍니다. |
“에이전트는 실수를 합니다. 때때로 당신이 완전히 옳다고 말하면서도 잘못된 행동을 하곤 합니다.” – Guillermo Rauch, Vercel CEO
AI 시대에 품질을 제공한다는 것은 그 토큰이 실제로 생산하는 것의 기준을 높이는 것을 의미하며, 단순히 토큰 수를 세는 것이 아닙니다. 스킬은 에이전트에 영향을 미치고, 프레임워크 변화에 맞춰 최신 상태를 유지하며, 올바른 답변으로 바로 연결되는 경로를 제공함으로써 토큰 효율성을 높이는 방법을 제공합니다.
핵심 요약
- 2 million installs는 실제 채택 신호입니다.
- Security partnerships는 팀에게 신뢰할 수 있는 생태계를 만듭니다.
- 가장 흥미로운 작업은 CLI 수준이 아니라, 이제 skills를 대규모 지식 배포를 위한 일급 원시 요소로 다루는 에이전트와 도구에 있습니다.
우리는 계속 개발해 나갈 것입니다. 다음 Skills Night에 Come find us 혹은 skills.sh에서 만나세요.
.skills.sh에서
컨텍스트
- 이것이 나온 배경
- 우리가 해결해야 했던 보안 문제
- 데모가 보여준 내용
- 우리가 목표로 하는 것
솔루션
Socket – 은 교차‑생태계 정적 분석을 LLM 기반 잡음 감소와 결합하여 수행하며, 벤치마크 전반에 걸쳐 정밀도 95 %, 재현율 98 %, **F1 97 %**를 보고합니다.
Gen – 은 Sage라는 실시간 에이전트 신뢰 계층을 구축하고 있습니다. 이 계층은 에이전트의 모든 입·출력 연결을 모니터링하여 데이터 유출이나 프롬프트 인젝션 위험 없이 자유롭게 실행될 수 있게 합니다.
Snyk – 은 패키지‑보안 분야의 경험을 skills 컨텍스트에 적용하고 있습니다.