Skills vs slash commands vs agents

Slash commands

The narrowest primitive. A slash command is a shortcut you type — /help, /clear, /config. Built-in slash commands come with Claude Code and do things like clearing the transcript or opening settings. They are not extensible; they are the tool’s own UI.

Most user-authored “commands” are actually skills. Which brings us to…

Skills

A named, invokable prompt shape. Lives in a directory. Has a manifest, a prompt body, optional attached files. Runs in the current session. Zero cross-session state by default.

Reach for a skill when:

• The workflow is repeatable, self-contained, and worth codifying.
• You want it available as /name inside a session.
• You want teammates to have it too (project-local skills ship in the repo).

Subagents

A fresh Claude with its own context and its own tools. Runs in an isolated conversation. Returns one summary message.

Reach for a subagent when:

• The task is broad enough that its raw output would clutter your main context.
• You want isolation — the sub-work should not leak into the main thread.
• You want to spawn several in parallel to compress wall-clock time.

The decision tree

1. Is it a built-in UI action? (Clear the screen, view help, open settings.) → Slash command. You do not build these; they exist.
2. Is it a repeated prompt shape you want to invoke in the current session? → Skill.
3. Is it a large sub-task you want to hand off to a fresh instance and get a summary back? → Subagent.

Skills and subagents can compose. A skill can spawn a subagent. A subagent can, in principle, invoke a skill in its own session.

Common confusions

• “My skill is slow because it does a lot.” → It may want to spawn a subagent for the heavy step. Skill on the outside, subagent inside.
• “My subagent keeps making the same mistake.” → The briefing may want to become a skill so the mistake gets fixed once, not every time you spawn.
• “I only need this once.” → Neither a skill nor a subagent. Just write it as a prompt in the current session.

Check your understanding

Quick check

—

Your team runs the same "audit these 5 files for X" workflow constantly, and each run explores a lot of files whose raw output you do not want cluttering the main session. What's the right architecture?

1. AA slash command that clears the transcript first.
2. BA skill that briefs and spawns a subagent — captures the workflow (skill) and isolates the exploration (subagent).
3. CNothing — repetition is fine.
4. DA subagent alone — no skill needed.

Check answer→