How to Use AMD Skills with Claude Code, Cursor and Codex

01

What AMD Skills is

AMD Skills is AMD’s official catalog of focused agent workflows. Each skill centers on a SKILL.md file and can include instructions, scripts, templates and references that tell an agent how to perform an AMD-specific task.

AMD positions Skills as one part of ROCm.AI alongside ROCm CLI and Hyperloom. For the broader release context, see what changed in ROCm 10.

02

The catalog is live, but it is still a Tech Preview

AMD’s August 27 release says shipped skills are available through the open GitHub catalog and supported agent marketplaces. The repository itself is explicitly labeled Tech Preview and warns that skills, categories and descriptions can change frequently.

03

Which coding agents are supported?

The official AMD Skills repository names Cursor, Claude Code, OpenAI Codex and Gemini CLI as interoperable coding-agent environments. AMD’s ROCm 10 launch specifically highlights Claude Code, Cursor and Codex for AMD-specific guidance.

Agent-specific install destinations differ, so follow the live repository instructions instead of assuming one plugin directory works everywhere.

04

Install from AMD’s current catalog

AMD currently documents installation with the skills CLI: npx skills add amd/skills. The command lets you select a skill and destination, and the repository also documents flags for installing a chosen skill into agents such as Claude Code, Cursor or Codex.

05

Use skills for outcomes, then inspect the actions

A skill can help an agent carry out an outcome-oriented task such as setting up local AI, serving a model or diagnosing an AMD environment without requiring the user to remember every underlying command.

The skill does not override the ROCm compatibility matrix. Before an agent changes drivers, system packages or production services, verify the target GPU and operating-system path and review the generated commands.

06

How Skills relates to ROCm CLI and Hyperloom

Skills provides higher-level task knowledge for agents. ROCm CLI is AMD’s Technology Preview execution layer for environment and runtime operations. Hyperloom is a narrower agentic optimization system for end-to-end LLM inference. A workflow can combine them, but each component has its own maturity and compatibility constraints.

07

A safe practical workflow

1. Confirm the hardware and ROCm path are supported
2. Browse the current AMD Skills catalog
3. Install the specific skill into the intended coding agent
4. Ask for the outcome you want in plain language
5. Review commands that modify system or production state
6. Validate the result with ROCm and framework health checks
7. Record the skill and source revision for reproducible production workflows

For the underlying platform setup, use the ROCm 10 compatibility and installation guide rather than assuming an agent skill can make unsupported hardware work.