reusable context and instructions for AI agents — loaded on demand, not crammed into prompts.
skills are packages of instructions, reference docs, and scripts that agents retrieve when relevant. instead of a massive system prompt that slows every response, your agent sees a short list of available skills and loads what it needs.
how skills work
when you add skills to an agent:
- the agent's
skill_gettool lists available skills (name + description) - when relevant, the agent calls
skill_get("skill-name")to load the full content - supporting files (references, scripts, examples) are loaded on demand via
skill_get("skill-name", "references/api-docs.md")
this progressive disclosure pattern keeps the context window focused. the same approach is used by Claude Code, Cursor, and other leading agents.
1user: "review this pull request"23agent: [calls skill_get(skill="code-review")]4 → loads code review guidelines56agent: "based on our code review guidelines, here's my analysis..."the agent skills standard
skills follow the agent skills open standard — a shared format that works across tools. a skill is a directory with a SKILL.md entry point plus optional supporting files:
1my-skill/2├── SKILL.md # entry point — yaml frontmatter + markdown instructions3├── references/ # domain docs, API specs — loaded on demand4│ └── api-docs.md5├── scripts/ # executables the agent can run6│ └── helper.py7├── examples/ # example outputs showing expected format8│ └── sample.md9└── templates/ # templates for the agent to fill in10 └── template.mdinference.sh implements this standard natively — skills you create here work on inference.sh agents, Claude Code, Cursor, Windsurf, and any tool that supports the spec.
two ways to consume skills
| inference.sh agents | local agents (Claude Code, Cursor, etc.) | |
|---|---|---|
| how it works | agent calls skill_get → content fetched from URL at runtime | skill directory on local filesystem, agent discovers by scanning |
| loading | on-demand via HTTP | on-demand from disk |
| no setup needed | just add the skill to your agent config | install the skill directory to ~/.claude/skills/ |
| paths work everywhere | skill_get("my-skill", "references/api-docs.md") | relative path from SKILL.md: references/api-docs.md |
the same paths work in both models. SKILL.md says references/api-docs.md — that works as a skill_get argument against our API and as a relative filesystem path locally.
what's next
- creating skills — build and publish skills on inference.sh
- the registry — browse, search, and share public skills
- adding skills to agents — attach skills to your inference.sh agents
- using with other agents — install skills to Claude Code, Cursor, and more