Built-in Slash Commands Tour

Your prompt input speaks two languages. Most of the time you type natural-language instructions and Claude does the work. But when a line starts with /, Claude Code intercepts it first and treats it as a slash command — a direct control over the session itself: switch models, clear context, run a review, open a manager.

Three parsing rules govern every one of them:

1. Recognized only at the start of a message. /clear on its own line clears context. please run /clear for me is just text Claude reads — the /clear in the middle does nothing.
2. Trailing text becomes arguments. Everything after the command name is passed to the command. /compact focus on the auth refactor runs /compact with focus on the auth refactor as its instruction.
3. Type / to discover, /help to be sure. Typing / opens an autocomplete menu of every command available to you right now; add letters to filter. Because the set changes with your version, your plugins, and your connected MCP servers, /help is the authoritative list for your install — this lesson teaches the families and the high-value commands, not a frozen catalog.

Think of slash commands as the console for the session, sitting alongside the conversation. Same input box, two distinct languages.

Not every slash command is the same kind of thing. The commands page tags each entry, and the tag tells you who can trigger it and how it behaves.

The crucial consequence is in that last column. A bundled skill works exactly like a skill you write yourself: it's a prompt that Claude can also invoke autonomously when the moment calls for it. So /code-review isn't only something you type — if you ask Claude to "check my diff for bugs," it may run the same review skill itself. /claude-api likewise activates on its own when your code imports anthropic or @anthropic-ai/sdk.

Plain built-ins (/clear, /model, /effort, /usage) are pure session controls — Claude never runs those for you; they only do something when you type them.

These manage the conversation itself: its size, its history, and where you are in it. This is the family you'll touch most, because context management is the core skill of working with Claude Code.

The everyday rhythm: /clear between unrelated tasks so context from one doesn't pollute the next; /compact when a single long task fills the window; /context to see why it's full before you decide which to use.

Model & effort controls how much brain you spend on the next turn — a direct lever on cost, latency, and depth.

The effort levels, lowest to highest, are: low, medium, high, xhigh, max, ultracode. Which levels are available depends on the model. Two are session-only (you can't save them as a default): max and ultracode. And ultracode is special — it's a Claude Code setting that combines xhigh reasoning with automatic workflow orchestration, so Claude not only thinks harder but can fan the work out across subagents on its own. /effort auto resets to the model's default.

Review & quality are bundled skills (the [Skill] tag) — meaning Claude can also run them autonomously — that audit your current git diff.

The division of labor is clean: /code-review finds bugs, /simplify cleans up, /security-review checks security.

Workflow orchestration scales work out across plan mode, subagents, and the background. Several carry the [Skill] or [Workflow] tag, so Claude can also reach for them itself.

Utilities & diagnostics are the supporting cast — copying output, exporting, checking spend, and fixing the install.

Keep /help, /usage, and /doctor in muscle memory: one tells you what's available, one tells you what you're spending, one tells you what's broken.

Two once-common commands are gone, and following stale tutorials that still list them is a common source of confusion:

If you type a removed command on a current version, it simply isn't recognized — which is another reason /help (not a blog post) is your source of truth.

Finally, your command surface isn't fixed — it grows with your connected MCP servers. An MCP server can expose prompts, and each surfaces as a command in the format /mcp__<server>__<prompt>: double-underscore separated, namespaced by server. These are dynamically discovered from your connected servers, so they appear in your / autocomplete only when the relevant server is connected. For example, a connected linear server exposing a create_issue prompt would surface as /mcp__linear__create_issue.