Agent Skills

Packaged, reusable capabilities Claude loads on demand. Each skill is a directory with a SKILL.md (instructions + frontmatter) plus optional scripts and resources.

Skill anatomy

my-skill/
  SKILL.md            ← required: instructions + description frontmatter
  scripts/            ← optional helper scripts
  templates/          ← optional reference files
  ...

SKILL.md frontmatter:

---
name: pptx
description: Use whenever a .pptx file is involved — creating, reading, editing slide decks.
---

The description is critical: Claude reads it to decide when to invoke the skill. Treat it like a tool description — be specific about triggers AND non-triggers.

How skills are loaded

On session start, Claude scans available skills (their names + descriptions).
During the conversation, when relevance is detected, the skill body is loaded into context.
Skills are lazy: only the description is in context until invoked. Cheap to have many.

Locations

~/.claude/skills/ — user-global.
<repo>/.claude/skills/ — project-shared (committed).
Bundled in plugins.

Skills vs slash commands vs subagents

	Slash command	Skill	Subagent
Invoked by	User typing `/x`	Model (auto-detected)	Model (delegating a task)
Context cost when idle	None	Just the description	None
Can include files/binaries	No	Yes	N/A (it's an agent)
Best for	Quick reusable prompt	Reusable capability (e.g., "make a PowerPoint")	Isolated focused task

Writing a good skill

Description is the prompt that gets Claude to call it. Be a teacher.
Lead with concrete triggers ("Use when the user mentions a .pptx file").
Include negative triggers ("Do NOT use for PDFs — see pdf skill").
Keep SKILL.md body short. Push details into referenced files Claude can read on demand.
For language-specific scripts, include them in scripts/ and reference exact relative paths.

Distribution

Plugins (the .plugin format) can bundle multiple skills.
Enterprise can deploy skills via managed config.

CCA-F angle

Distinguish:

Skill = capability bundle.
Slash command = parameterized prompt template.
MCP prompt = server-provided prompt template.
Subagent = isolated agent.

These often confuse exam-takers. Memorize who invokes (user vs model vs server) and what gets included (just text vs files vs separate context).