Slash Commands

Verified on 2026-10-04 with Claude Code 2.1.289.

Slash commands are how you steer a session without writing a prompt: switch models, free up context, branch a conversation, review a diff, or hand a long task to a loop that keeps going until a condition holds. This chapter covers the commands that matter most to a developer who runs agents all day, and the few that changed meaning since early 2026.

The full list is long and changes almost weekly. Treat the official commands reference as the source of truth, and this chapter as the map of what to reach for and when.

How commands work

Type / at the prompt to open a filterable menu, then keep typing to narrow it. A few things are worth knowing about how the menu and the parser behave:

• Only at the start. A command is recognized only at the start of your message. Anything after the name becomes its arguments.
• Queued or immediate. If you send a command while Claude is still responding, Claude Code usually queues it until the turn ends. Some, such as /status, /tasks and /usage, run immediately without interrupting the response.
• Not everything shows for everyone. Availability depends on platform, plan and provider. /desktop, for example, appears only on macOS and x64 Windows with a Claude subscription, and the artifact-based commands are unavailable on Amazon Bedrock, Google Cloud and Microsoft Foundry.
• Built-ins vs bundled skills. Most entries are built-in commands with fixed behavior. Others, marked Skill in the reference, are bundled skills: prompts handed to Claude, which then does the work with its tools. /code-review, /doctor, /loop, /batch, /design and /verify are bundled skills. The difference matters: a bundled skill spends tokens and its quality depends on the model and effort level.
• Your skills share the menu. Skills you write appear alongside built-ins, and MCP servers can expose prompts that appear as commands. A skill with the same name as a built-in replaces it in a local terminal session, though not its aliases. See Custom Skills.

The commands you will use daily

Context and conversation

Command	What it does
/clear [name]	Starts a new conversation with empty context. Pass a name to label the old one in the /resume picker. Aliases: /reset, /new
/compact [instructions]	Summarizes the conversation to free space, optionally focused by your instructions
/context [all]	Shows context usage as a colored grid, with suggestions for heavy tools, memory bloat and capacity warnings
/resume [session]	Resumes a conversation by ID or name, or opens the picker. Alias: /continue
/rename [name]	Names the session; without a name, generates one from the conversation
/rewind	Rolls code and conversation back to a checkpoint, or summarizes from a chosen message. Aliases: /checkpoint, /undo
/btw [question]	Asks a side question without adding it to the conversation history
/export [filename]	Exports the conversation as plain text

/clear is no longer a one-way door. In the same Claude Code process you can get the cleared conversation back from the rewind menu, or with /resume.

Model, effort and modes

Command	What it does
/model [model]	Switches the model and saves it as your default for new sessions. In the picker, press s on a row to switch for this session only
/effort [level|auto|status]	Sets effort: low through xhigh, max, or auto. max lasts for the session only
/plan [description]	Enters plan mode; with a description, starts planning that task right away
/fast [on|off]	Toggles fast mode
/permissions	Manages allow, ask and deny rules, and shows recent auto mode denials. Alias: /allowed-tools
/config [key=value ...]	Opens settings, or sets a key directly, for example /config theme=dark

Two behaviors are easy to miss. /model now persists your choice as the default unless you use the session-only s key. And changing the model or effort mid-session can cost you the prompt cache: Claude Code warns before a switch that would invalidate it. Voice, Fast Mode and Effort and Tokens, Limits and Caching cover the trade-off.

Before you ship

Command	What it does
/diff	Reviews working-tree changes, including Claude's edits so far
/code-review [level] [--fix] [--comment] [--max-findings n] [target]	Reviews the current diff, or a PR number, branch or path, for correctness bugs. Alias: /review
/security-review	Checks the branch's diff against origin's default branch for security vulnerabilities
/simplify [target]	Looks for cleanup opportunities (reuse, simplification, efficiency, abstraction level) and applies them. It does not hunt for bugs
/verify	Builds and runs your app to confirm a change works, rather than relying on tests alone. Runs only when you invoke it

The new commands in depth

/goal: keep working until a condition holds

/goal sets a completion condition and lets Claude keep taking turns without you prompting each one. After every turn, a separate model checks whether the condition holds. If not, Claude starts another turn. The goal clears when the condition is met, when the evaluator judges it impossible, when a turn fails on an error you have to fix, or when you clear it.

/goal all tests in test/auth pass and the lint step is clean

Setting a goal starts a turn at once, with the condition itself as the instruction. Run /goal with no argument to see the condition, how long it has run and how many turns were evaluated. /goal clear (or stop, off, reset, none, cancel) ends it early. One goal can be active per session.

Three things make a goal work well:

1. The evaluator reads the transcript, not your disk. It does not run commands or open files itself. Write the condition so that Claude's own output proves it: "npm test exits 0" works because the test output lands in the conversation.
2. One measurable end state, plus constraints. A test result, a build exit code, a file count, an empty queue. Add what must not change, such as "no other test file is modified".
3. A bound. Add a clause such as or stop after 20 turns. The condition can be up to 4,000 characters.

A goal does not change your permission mode. In manual mode Claude still stops to ask before unapproved tool calls, so a goal only runs unattended in auto mode. The docs describe the pairing this way: auto mode removes per-tool prompts, and /goal removes per-turn prompts.

/goal is the session-scoped cousin of a Stop hook. A Stop hook lives in settings and applies to every session in its scope; /goal is typed once and lasts for this session. /loop, by contrast, repeats on a time interval. Orchestrating Many Agents shows /goal inside larger setups.

/branch, /fork and /subtask: three ways to split a conversation

These three names are easy to mix up, and their meanings shifted during 2026. As of 2.1.289:

Command	What you get	Where the result goes
/branch [name]	A copy of the conversation, and you switch into it. The original stays available through /resume	You keep working in the branch
/fork [prompt]	A copy of the conversation as a new background session, while you keep working here	The copy runs on its own; watch it in claude agents
/subtask <task>	A forked subagent that inherits the full conversation and works in the background	Its result comes back into this conversation

/fork with a prompt starts the copy working on it immediately. Unless the copy edits in place, Claude Code tells it to create its own worktree before changing code, so two copies do not trample each other's files. On versions 2.1.161 through 2.1.211, and whenever agent view is turned off, /fork starts a forked subagent instead, which is what /subtask does today. If an older tutorial says /fork "reports back", that is the behavior it describes.

Use /branch to try a different direction yourself, /fork to let a second session pursue an idea in parallel, and /subtask for a side task whose answer you need in this conversation. Subagents and The Agent View and Sessions go deeper.

/usage: one place for cost, limits and stats

/usage shows session cost, your plan's usage limits and activity stats. On Pro, Max, Team and Enterprise plans it includes a breakdown of what counts against your limits, including which skills, subagents and MCP servers drove usage. /cost and /stats are now aliases of /usage (/stats opens on the Stats tab).

When a usage limit blocks you, /rate-limit-options lists ways to keep going: wait and continue automatically when the limit resets, add usage credits, or upgrade. Tokens, Limits and Caching covers how to read these numbers.

/doctor: setup checkup and prompt audit

/doctor used to be a read-only diagnostics screen. It is now a bundled skill that checks installation health (duplicate installs, PATH problems, unparseable settings files), flags slow hooks, finds skills, MCP servers and plugins whose context cost outweighs their use, and can trim checked-in CLAUDE.md files by cutting what Claude could derive from the code. It reports findings first and asks before changing anything. Alias: /checkup.

/doctor
/doctor prompt-audit

/doctor prompt-audit (2.1.283 and later) is a different job: Claude audits your CLAUDE.md files, skills, agents and commands for outdated or conflicting instructions, including prompting patterns written for older models. Run it after a model upgrade. From the shell, claude doctor prints read-only installation diagnostics without starting a session.

/code-review: review at a chosen depth

/code-review reviews the current diff for correctness bugs, or a PR number, branch or path you pass. It takes an effort level and a few flags:

/code-review
/code-review high 1234
/code-review --fix
/code-review --comment
/code-review ultra

• The level (low, medium, high, xhigh, max) sets depth. With no level, it reuses the last one you typed.
• --fix applies the findings. --comment posts them on the GitHub PR or GitLab merge request.
• --max-findings <n>|all|default (2.1.288 and later) reports more or fewer findings than the usual limit.
• ultra runs a deep multi-agent review in a cloud sandbox (ultrareview). It includes 3 free runs on Pro and Max, then needs usage credits. /ultrareview is an alias, and claude ultrareview runs it from the shell.

/code-review runs as a background subagent, so you can keep working while it reviews. /review is now an alias of /code-review; before 2.1.223 it was a separate single-pass PR reviewer.

/design: draft UI as artboards

/design [brief] (research preview, 2.1.265 and later) drafts UI mockups, screen flows, landing pages or posters as artboards on one canvas, published as a Claude Design artifact. You edit the artboards in a desktop browser, export them as PNG or PDF, and have Claude implement the one you pick.

/design a settings screen for a mobile banking app

It needs a session where artifacts are available and an account where the Design template is on, so it is unavailable on Bedrock, Google Cloud, Foundry and Claude Platform on AWS. Agents That See covers the visual loop.

Other commands worth knowing

Command	What it does
/background [prompt]	Detaches this session to run as a background agent and frees the terminal. Alias: /bg
/tasks	Lists background work in this session, including finished subagents
/batch <instruction>	Splits a large change into 5 to 30 independent units, each in its own worktree and background subagent
/loop [interval] [prompt]	Repeats a prompt while the session is open; omit the interval and Claude paces itself
/schedule [description]	Creates and manages cloud routines. Alias: /routines
/skills, /skill-doctor, /reload-skills	List skills, show each skill's context cost and usage, and pick up skills changed on disk
/plugin, /reload-plugins	Manage plugins and apply changes without restarting
/hooks, /mcp	Show hook configuration; manage MCP servers (/mcp reconnect all retries failed ones)
/memory, /init	Edit CLAUDE.md and auto memory; generate a starter CLAUDE.md
/import [codex|gemini|cursor]	Brings instruction files, MCP servers, commands, subagents and skills over from another agent. --dry-run previews
/advisor [model|off]	Pairs your model with a stronger advisor it consults at key moments
/autocompact [auto|<tokens>]	Sets how full the context gets before auto-compaction
/output-style [style]	Lists or switches output styles, for example concise
/remote-control, /teleport, /desktop	Continue this session from another device, pull a cloud session here, or open it in the desktop app
/status, /help, /release-notes	Version, model and account; help; changelog picker

Removed or renamed commands

• /pr-comments was removed in 2.1.91. Ask Claude to read the PR comments instead.
• /vim was removed in 2.1.92. Use /config and set the editor mode.
• /agents no longer opens an editor (since 2.1.198). It reminds you to ask Claude to create subagents or to edit .claude/agents/ directly.
• /mobile shows a QR code to download the mobile app; it does not change the layout.
• /ultraplan was removed. Use plan mode.

Patterns that pay off

Pair /goal with a check Claude can show. "Done" should be a command whose output lands in the transcript. Vague goals ("make it better") run until the evaluator gives up or your turn bound hits.

Run /context before you /compact. It tells you what is actually filling the window. Often the answer is an MCP server or a skill listing, not the conversation, and compaction will not fix that.

Review at two depths. Use /code-review at your normal level on every change, and save ultra for changes that touch money, auth or data. Use /simplify separately; it looks at cleanup, not bugs.

Use /btw for questions that should not stick. It answers from current context and leaves no trace in history, so it does not crowd out the task.

Run /doctor prompt-audit after every model change. Instructions written to coax an older model can fight a newer one.

Check that it worked

1. Run /status and confirm the version line shows 2.1.289 or later. Several commands in this chapter have minimum versions.
2. Type /go and confirm /goal is highlighted in the menu. If it is missing, see Troubleshooting.
3. Run /usage and confirm you see your plan's limits (or session cost on an API key).
4. Run /doctor. It should finish with a findings report and ask before changing anything.

Sources

• Commands reference, Anthropic, accessed 2026-10-04. https://code.claude.com/docs/en/commands
• Keep Claude working toward a goal, Anthropic, accessed 2026-10-04. https://code.claude.com/docs/en/goal
• What's new, weekly digests for Weeks 20–37 (May 11 – September 11, 2026), Anthropic. https://code.claude.com/docs/en/whats-new
• Claude Code changelog, entries 2.1.283 (2026-09-25) and 2.1.288 (2026-10-02), Anthropic. https://code.claude.com/docs/en/changelog
• claude --help, Claude Code 2.1.289, run 2026-10-04.