Agent Skills
Modular, on-demand capabilities that an AI agent loads only when relevant to the current task—each skill bundles a short instruction file, optional scripts, and reference documents. Skills let one agent serve many domains (refund policy, contract review, SEO audit) without bloating the system prompt: the agent reads a small skill index, decides which skill applies, and loads its full instructions only when needed. Popularized by Anthropic's Claude Skills and adopted in Claude Code, agent platforms, and enterprise deployments.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A support agent has 40 skills installed: refund-policy, shipping-eta, gdpr-deletion, account-merge, b2b-procurement, and so on. The system prompt itself is only 800 tokens and lists each skill's name and one-line trigger. When a customer asks about a refund, the agent loads the full refund-policy skill (1,200 tokens of policy detail and edge cases), executes it, and unloads—keeping the working context lean even though the agent's total knowledge spans 50,000 tokens of skill content.
Frequently asked questions
- How are skills different from RAG over a knowledge base?
- RAG retrieves passages of reference content that the model reads. Skills retrieve executable instructions plus optional scripts and resources that the model follows. A RAG result tells the agent what is true; a skill tells the agent what to do—including which tools to call and in what order. The two compose well: a skill often calls RAG as one of its steps.