Claude Code Skills & Slash Commands: When to Use Which (CCA-F)

Claude Code gives you four ways to teach an agent something, and the CCA-F exam's favorite Configuration-domain question is simply: which one? CLAUDE.md, skills, slash commands, and MCP servers overlap enough to confuse, and the wrong choice is a real cost (a bloated context, a manual step that should have been automatic, a capability that never loads when needed). This guide draws the lines.

Skills and slash commands live in the Claude Code Configuration & Workflows domain (20% of the exam). For the file that anchors persistent memory, pair this with the CLAUDE.md guide; for the enforcement layer, the hooks guide.

Skills: Capabilities That Load on Demand

A skill is a packaged capability the agent pulls in when it is relevant, not every session. Structurally, a skill is a folder with a SKILL.md (instructions plus a description of when to use it) and optionally supporting files, scripts, or resources. The key property is progressive disclosure: only the skill's short description sits in context normally; the full instructions load when the agent decides the skill applies.

That solves the CLAUDE.md bloat problem directly. A 300-line deployment runbook does not belong in memory that loads every session. As a skill, its one-line description ("use when deploying to production") costs almost nothing until the moment a deploy comes up, at which point the full procedure loads.

Skills can be personal (~/.claude/skills/), project-scoped (.claude/skills/, shared via git), or provided by plugins. The description field is doing the same routing job it does for subagents: it is how the agent decides whether to load the skill at all, so vague descriptions leave skills stranded.

Slash Commands: User-Triggered Shortcuts

A slash command is a reusable prompt the user invokes by typing /name. It is a Markdown file in .claude/commands/ (project) or ~/.claude/commands/ (personal) whose body is the prompt that gets sent, optionally with $ARGUMENTS interpolation.

The defining difference from a skill: who decides to use it. A skill is loaded by the agent when it judges the skill relevant. A slash command fires when you type it. /review-pr 1234 is you deciding to run a saved workflow; a skill is the agent deciding, mid-task, that it needs the PDF-processing capability.

Slash commands shine for workflows you run repeatedly and want to trigger deterministically: /deploy, /write-tests, /changelog. They are muscle memory, not agent judgment.

The Full Selection Matrix

This table is the exam's Configuration domain in one view:

The sorting questions the exam asks, and their answers:

• "A 5-step incident-response procedure used maybe twice a month" -> skill (on-demand, keeps context lean)
• "The team's standard PR-review prompt, run several times a day" -> slash command (user-triggered, repeatable)
• "The repo's test command" -> CLAUDE.md (needed every session)
• "Block writes to the secrets file" -> hook (hard guarantee)
• "Let the agent query the production database" -> MCP server (external system access)

Skills vs MCP: The Subtle One

Both extend what an agent can do, and professional questions probe the boundary. The clean distinction:

• MCP connects the agent to external systems: a database, a SaaS API, a file store. It is about access and integration, exposing tools/resources/prompts over a protocol.
• Skills package knowledge and procedure: how to do a task well, what steps in what order, which files to touch. A skill can use MCP tools, but the skill itself is instructions, not a connection.

A "process customer refunds" skill might describe the policy and steps; the MCP server it calls is what actually reaches the billing system. Confusing the two ("use a skill to connect to the CRM," "use MCP to teach the refund procedure") is a classic wrong answer. See the MCP architecture guide for the integration half.

Sharing and Team Workflows

A quietly important exam theme: which of these travel with the repo. Project-scoped skills, slash commands, and CLAUDE.md are committed to git, so cloning the repo gives every teammate (and every agent instance) the same capabilities and conventions. Personal-scoped versions stay on your machine. Designing a team's Claude Code setup means deciding deliberately what is shared project config versus personal preference, and the exam rewards putting shared conventions in project scope.

Key Takeaways

• Skill = agent loads it on demand (progressive disclosure); slash command = user triggers it by typing
• CLAUDE.md = every session; skill = some sessions; hook = must happen; MCP = external access
• Skills package procedure and knowledge; MCP packages connection to systems; a skill may call MCP tools
• Project-scoped skills/commands/CLAUDE.md travel via git; put shared conventions there
• Descriptions are routing logic for both skills and subagents, so write them precisely

The Claude Code course builds real skills and slash commands in Level 5, and Preporato's CCA-F practice tests drill the selection matrix across 390 explained questions.

Sources:

• Claude Code Documentation: Skills
• Claude Code Documentation: Slash Commands
• Anthropic Academy

Last updated: July 10, 2026