Agent Skills: Procedural Knowledge Packets

Agent skills are dynamic units of procedural knowledge stored as folders with a SKILL.md file. They load on demand, avoiding the need to bake every capability into model weights. This approach reduces computational waste and enables reusable workflows.

SKILL.md Structure

The SKILL.md file has a YAML frontmatter with a Name and Description. The body contains step-by-step instructions. Optional subdirectories can include scripts or references.

• YAML Frontmatter: Mandatory for metadata.
• The Body: Holds execution logic.
• scripts/: For executable code.
• references/: For supporting documents.

Progressive Disclosure

Agents load skills in stages to manage context:

1. Metadata only at startup.
2. Full instructions on task match.
3. Scripts and assets when needed.

This prevents context window saturation, especially with hundreds of skills.

Comparison with MCP and RAG

Component	Purpose
MCP	External API interaction
RAG	Factual data retrieval
Skills	Procedural task logic

Skills provide the how-to for tasks, unlike MCP's tool access or RAG's data supply.

Conclusion

Agent skills offer a scalable way to add capabilities without model retraining. They promote modularity and efficiency in AI agent design.