** AI 코딩 에이전트에 대한 공식 Intel-maintained 전문가 워크플로우. **

빌드, 포트, 실행, 프로필 및 인텔 플랫폼에서 소프트웨어 최적화. 각 기술 캡처 제품 문서가 아닌 실제 작업 흐름: 그것은 환경을 검사, 선택 지원되는 경로, 일은, outcome를 검증하고, 실패를 처리한다. 실제로 일어나.

현재 적용 범위: 인텔 GPU 설정 및 진단, AI 런타임 및 모델 서빙, 배포 계획 및 벤치마킹, CUDA-to-XPU 마이그레이션, CPU 성능 분석, Python Intel 장치의 가속, OneTBB 개발.

카탈로그를 찾아 → · 기여 · 유효성 검사

퀵스타트

이 저장소를 직접 읽는 표준 기술 CLI로 설치하십시오:

그것은 카탈로그를 열고 당신이 선택할 수 있습니다. 아는 세 가지 변형 :

설치는 기본적으로 Project-scoped - Claude Code의 ./.claude/skills/ 및 각 다른 에이전트에 대 한 동등한 디렉토리 이름. --global per-user 위치에 쓰기 대신. 읽는 디렉토리에 대한 에이전트의 자신의 문서를 확인합니다.

pinned, 오프라인, byte-verifiable install path도 존재합니다. Reproducible 및 오프라인 설치.

Agent Skills

Official Intel-maintained expert workflows for AI coding agents.

Build, port, run, profile and optimize software on Intel platforms. Each skill captures a practical workflow rather than product documentation: it inspects the environment, chooses a supported path, does the work, verifies the outcome, and handles the failures that actually happen.

Current coverage: Intel GPU setup and diagnosis, AI runtimes and model serving, deployment planning and benchmarking, CUDA-to-XPU migration, CPU performance analysis, Python acceleration on Intel devices, and oneTBB development.

Browse the catalog → · Contribute · How validation works

Quickstart

Install with the standard skills CLI, which reads this repository directly:

npx skills add intel/skills

That opens the catalog and lets you pick. Three variants worth knowing:

# browse without installing anything
npx skills add intel/skills --list

# one skill, one agent
npx skills add intel/skills --skill vllm-xpu-run --agent claude-code

# for every project on the machine rather than this one
npx skills add intel/skills --skill linux-perf --agent claude-code --global

Installs are project-scoped by default — ./.claude/skills/ for Claude Code, and the equivalent directory for each other agent you name. --global writes the per-user location instead. Check your agent’s own documentation for the directory it reads.

A pinned, offline, byte-verifiable install path also exists, for controlled environments: Reproducible and offline installation.

Using what you installed

Describe the task in your own words and start a new session. The agent reads each installed skill’s description and opens the one that matches, so naming it is optional:

“Serve Qwen/Qwen3-8B on my Intel GPU with an OpenAI-compatible endpoint.”

Naming it works too, and is the fastest way to check an install took:

“Use the vllm-xpu-run skill to serve Qwen/Qwen3-8B.”

npx skills list              # what is installed, where, and for which agent
npx skills update            # pull the current version of each installed skill
npx skills remove            # take one back out

If the agent does something generic and never opens a skill you know covers the task, name the skill explicitly to confirm it is installed and being read — then open an issue with the wording you used. A skill an agent does not reach for is a defect in its description, it fails silently, and the request you actually typed is the only thing that finds it.

Which skill covers what: intel.github.io/skills, searchable and filterable by product and hardware, generated from this repository on every push. Several skills chain — a deployment plan calls preflight, sizing and configuration