---
id: "concept-agent-skills"
type: "concept"
source_timestamps: ["00:01:46", "00:03:42", "00:04:17"]
tags: ["prompt-engineering", "agent-context", "documentation"]
related: ["concept-claude-code", "concept-remotion"]
definition: "Machine-readable documentation and rule sets installed locally to teach AI agents how to correctly use specific frameworks or libraries."
---
# Agent Skills

## Definition

Machine-readable documentation and rule sets installed locally to teach AI agents how to correctly use specific frameworks or libraries.

## Structure

When a user installs an Agent Skill (e.g., `npx skills add remotion-dev/skills`), it downloads a directory containing:

- A `SKILL.md` file describing the skill at a high level
- Specific rule files codifying best practices and gotchas
- Domain-specific knowledge unique to the target framework

These files act as a **highly concentrated context window injection** that the agent reads when relevant.

## Why They Matter

By reading these files, [[concept-claude-code|Claude Code]] bypasses its training data limitations or hallucinations and writes syntactically correct, up-to-date code for the target framework. For the [[concept-remotion|Remotion]] skill, this includes rules on:

- Animation handling
- Audio integration
- Font management
- Composition structure

## Implicit Invocation

A key UX property is that Agent Skills are triggered implicitly. The user doesn't need to type a magic command; mentioning the target framework in natural language is sufficient. See [[quote-implicit-triggering]] for Sabrina Ramanov's framing of this behavior.

## Related

- [[action-install-remotion-skill]] — concrete install command
- [[concept-mcp]] — complementary mechanism: skills add knowledge, MCP adds external tool access