Summary of "What AI Agent Skills Are and How They Work"

Concise summary

This document explains what AI agent skills are, how they work, and key takeaways for using and authoring them. Skills encode procedural knowledge (step‑by‑step workflows, rules, examples, and judgment) that supplements an LLM’s semantic/factual reasoning.

What skills are

Skills provide procedural knowledge for AI agents — step‑by‑step workflows, rules, examples, and judgment about how to perform repeatable tasks (for example, a 47‑step compliant report).
They supplement LLMs’ factual reasoning (semantic knowledge) by encoding “how to do things” (procedural memory).

Skill format and content

A skill is a folder with a core file named skill.md (plain Markdown) and optional supporting files.

skill.md
- Contains YAML front matter. Two mandatory fields:
  - name — the skill name
  - description — the trigger condition (what should cause the agent to use the skill)
- Optional YAML fields can include author, version, etc.
- The body of skill.md contains the actual instructions/workflow, examples, and rules.
Optional subfolders:
- scripts — executable code (JS/Python/bash) the agent can run
- references — additional documents loaded on demand
- assets — static templates or data files

Example of the minimal YAML front matter (illustrative):

yaml name: "PDF Builder" description: "Use this when the user asks to extract or construct a PDF"

How skills are loaded and triggered (progressive disclosure)

Skills are loaded progressively to minimize token cost and scale to many skills. The LLM decides when a skill matches a task, so clear, accurate descriptions are crucial.

Tier 1 (startup)
- Load only metadata (name + description) for every installed skill.
- Tiny token cost; acts like a table of contents.
Tier 2 (on match)
- When an agent’s request matches a skill description, the full skill.md is pulled into context.
Tier 3 (on need)
- Scripts, references, and assets are loaded only when required.

Relation to other knowledge integration approaches

MCP (Model Context Protocol): provides tool/API access (what the agent can reach) but not when/how to use it.
RAG (Retrieval‑Augmented Generation): supplies factual/semantic knowledge from a knowledge base, not procedural steps.
Fine‑tuning: bakes behavior into model weights (permanent and costly to update).
Skills complement these approaches by encoding procedural knowledge and often calling out to MCP or RAG when needed.

Benefits and properties

Portable plain‑text files suitable for version control and easy updates.
Open standard (agent‑skills.io, Apache 2.0).
Interoperable: skills written to the spec work across platforms that adopt it.
Efficient token usage via progressive disclosure; scalable to many skills.

Security and trust considerations

Skills can run local scripts and access file systems, environment variables, and API keys — powerful but risky.
Public skills have been found to include issues such as prompt injection, tool poisoning, or hidden malware.
Recommendation: treat skill installation like any software dependency — review and audit before running locally.

Practical notes and examples

Example skill name: “PDF Builder” — the description should state the trigger (for example, “use this when the user asks to extract a PDF”).
Cognitive‑science analogy for knowledge mapping:
- Semantic memory → RAG / knowledge bases
- Episodic memory → conversational logs
- Procedural memory → skill files

Covered as a tutorial / explainer

The source video explains:

How to structure a skill
Why descriptions matter
How loading works (progressive disclosure)
How skills interact with MCP, RAG, and fine‑tuning
Important security practices

Main speakers / sources

Video narrator/presenter (unnamed in subtitles)
Referenced standards and sites: agent‑skills.io (Apache 2.0 spec)
Referenced platforms: OpenAI Codex and other major AI coding platforms (e.g., Anthropic Claude and similar systems) that have adopted the spec