Skip to main content

Module 5 — Built-in slash commands (2/2): quality, parallelism, automation and integrations

Module 4 covered the commands that drive session, context, model and permissions. This module handles the other half that makes Claude Code useful in a team: review, parallel work, automation, extensions and integrations with GitHub, Slack and Chrome. The commands.md table is broken here into eight families, with usage on Kiosque, our food-truck app, that finally opens serious pull requests.

Vocabulary reminder: the official table marks Skill the bundled skills (a prompt shipped with the CLI, that Claude can also invoke on its own, except /verify since v2.1.215) and Workflow the dynamic workflows (/deep-research for instance, that Claude no longer invokes on its own since v2.1.218). The other commands are built-ins: their behavior lives inside the CLI. The distinction decides scriptability, token cost, and how to disable the command.

a. Quality and review

Five commands cover reviewing, plus a few satellites that run the application to verify it actually works. /code-review is the center of gravity, /review is its exact alias since v2.1.223, and /ultrareview is the historical name kept around.

CommandEffects
/diffShows the working-tree changes, including edits Claude made during the turn.
/code-review [low|medium|high|xhigh|max|ultra] [--fix] [--comment] [target]Bundled skill. Reviews the current diff, or a PR number, branch, path. --fix applies the fixes, --comment posts on the PR, ultra triggers a multi-agent cloud review. Without a level, reuses the last level typed from low to max.
/review [...]Exact alias of /code-review since v2.1.223.
/security-reviewAnalyzes the current branch against origin's default-branch HEAD: injections, authentication mistakes, data exposure. Without origin, the failure is ambiguous argument.
/simplify [target]Bundled skill. Four parallel agents: reuse, simplification, efficiency, right abstraction level. Does not look for bugs: use /code-review for that.
/verifyBundled skill. Builds, runs and drives the application to verify that a change produces the expected effect. Since v2.1.215, Claude can no longer invoke it on its own.
/runBundled skill. Runs the application to observe a change in operation.
/run-skill-generatorBundled skill. Writes the skill that teaches /run and /verify how to build and start the application from a clean environment.
/ultrareview [PR|branch]Historical alias of /code-review ultra. Three runs free on Pro and Max, then usage credits.

The key point is the background behavior: /code-review runs by default as a forked subagent with its own context window, therefore without polluting the current conversation. It falls back to foreground when a review is already running, in -p or via the Agent SDK, and if CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1. On an ultra cloud review targeting github.com, --post preselects publishing remarks in the launch box (v2.1.227+).

b. Parallelism and background

Three parallelization surfaces coexist: subagents (delegates inside the current session, separate context, result that comes back into the conversation), background sessions (whole Claude sessions living outside the terminal, listed by claude agents), and dynamic workflows (scripts that distribute work across many subagents and cross-check results).

CommandEffects
/plan [description]Switches to plan mode immediately (see module 8).
/tasksLists the session's background work, including finished subagents. Alias: /bashes.
/background [prompt]Detaches the current session as a background session. Alias: /bg.
/fork [prompt]Copies the conversation into a new background session while this one keeps running. The copy creates a worktree before any edit (v2.1.221+).
/subtask <task>Opens a forked subagent that inherits the whole conversation; the result flows back into the conversation (v2.1.212+).
/batch <instruction>Bundled skill. Breaks a change into 5 to 30 independent units, launches a background subagent per unit in an isolated git worktree, each runs its tests and opens a PR. Requires a git repo.
/agentsSince v2.1.198, only prints a reminder: ask Claude or edit .claude/agents/. Distinct from claude agents (agent view).
/list-agentsLists subagents, agent-team peers and sessions reachable via cross-session messaging. Alias: /peers. v2.1.224+.
/workflowsOpens the progress view for dynamic workflows.
/stopStops the attached background session; transcript and worktree are kept.
Ultraplan has been retired

/ultraplan no longer exists. The documentation explicitly points to the local plan mode (Shift+Tab, /plan), which replaces the former send-to-cloud session.

c. Automation

Four commands keep a session useful without supervision. /loop repeats, /goal pursues a condition evaluated after each turn by a small fast model, /schedule creates a cloud routine independent of the session, /autofix-pr delegates PR watching to a web session.

CommandEffects
/loop [interval] [prompt]Bundled skill. Repeats a prompt as long as the session is open. Without an interval, Claude picks a delay between 1 minute and 1 hour. Without a prompt, runs the built-in maintenance prompt or a loop.md (.claude/loop.md or ~/.claude/loop.md). Alias: /proactive.
/goal [condition|clear]Sets a condition. After each turn, a small fast model judges whether it is reached, impossible or not yet. clear, stop, off, reset, none, cancel clear the goal. Does not change permission mode.
/schedule [description]Creates, updates, lists or triggers a cloud routine. Alias: /routines.
/autofix-pr [prompt]Opens a web session that watches the current branch's PR and pushes fixes when CI fails or a reviewer comments. Requires gh.
The cost of a forgotten goal

A /goal consumes tokens every turn, and the evaluator consumes some at every verdict. Recurring /loop tasks expire after 7 days, but a /goal stays active until it is satisfied, judged impossible or cleared. Always bound the condition ("or stop after 20 turns") or type /goal clear at end of day.

d. Extensions

Extensions separate a bare install from team tooling: skills, plugins, hooks and MCP servers. Modules 6, 9, 11 and 12 cover them in depth; here, the commands that drive them.

CommandEffects
/skillsList of available skills. t sorts by token cost, Space/Enter cycles visibility toward Claude and toward the / menu. Plugin skills, skills with disable-model-invocation: true or a skillOverrides entry in managed settings do not move.
/skill-doctorContext cost and usage frequency of each skill. v2.1.252+.
/reload-skillsRescans skills and commands directories on the fly.
/plugin [subcommand]Manages plugins. list, install, enable, disable act directly.
/reload-plugins [--force]Reloads active plugins. Warns and does nothing without --force if reloading would break the cache.
/mcp [reconnect <server>|enable|disable [<server>|all]]Manages MCP connections and OAuth. Also in -p (v2.1.205+).
/hooksHook configuration for tool events.

e. Integrations

These commands hook Claude into an IDE, GitHub, Slack, Chrome, a mobile device, the desktop app. Most open a browser or a configuration window.

CommandEffects
/ideManages IDE integrations and shows the link state.
/install-github-appInstalls the GitHub app on a repo, with an optional step for GitHub Actions workflows and secrets.
/install-slack-appInstalls the Slack app via OAuth.
/chromeConfigures Claude in Chrome.
/web-setupConnects the GitHub account to Claude Code on the web via gh.
/teleportResumes a web session in the current terminal. Alias: /tp.
/remote-controlMakes the session drivable from claude.ai. Alias: /rc.
/remote-envChooses the default environment for cloud agents.
/desktopResumes the session in the desktop app. Alias: /app. macOS and Windows x64.
/design-loginAuthorizes design-system access for /design-sync.
/design-sync [name]Bundled skill. Converts the repo's React design system and sends it to Claude Design.
/design [brief]Bundled skill. Generates a canvas of UI mockups published as an artifact. Unavailable on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, Claude Platform on AWS.
/artifactsLists owned or shared artifacts, so you can attach one to the session. v2.1.208+.

f. Bundled knowledge skills

Alongside task skills (/code-review, /simplify, /verify), the CLI ships three knowledge skills that preload a reference topic so Claude does not invent it.

CommandEffects
/claude-api [migrate|upgrade|managed-agents-onboard|prompt-audit|cost-optimize]Bundled skill. Loads the Claude API and Managed Agents reference; activates automatically when the code imports anthropic or @anthropic-ai/sdk. Subcommands migrate code, bump the SDK a major version, guide the creation of a Managed Agent, audit legacy prompts, or profile spend.
/dataviz [request]Bundled skill. Helps design charts and dashboards: colorblind-safe palette, accessibility rules. Requires v2.1.198+.
/deep-research <question>Bundled workflow. Fans out research across the web, cross-checks sources, produces a sourced report. Claude no longer invokes it on its own since v2.1.218.

g. Diagnostic and feedback

When something is off, seven commands document, diagnose or report.

CommandEffects
/debug [description]Bundled skill. Enables the debug log for the current session and analyzes the log. Off by default except claude --debug.
/bug [report]Opens the bug report: choose the portion of session sent, consent screen. When signed with Anthropic, the send goes to Anthropic; otherwise ~/.claude/feedback-bundles/ receives the archive. Alias: /share.
/feedback [report]Product feedback; same box as /bug. Without an argument, in sessions with Claude-drafted feedback, opens the queue of drafts written by Claude.
/heapdumpWrites a memory snapshot to ~/Desktop. Hidden from the menu. The .heapsnapshot contains the whole conversation and identifiers — never share it.
/setup-bedrockConfigures Amazon Bedrock. Hidden as long as CLAUDE_CODE_USE_BEDROCK=1 is not set.
/setup-vertexConfigures Google Cloud's Agent Platform. Hidden as long as CLAUDE_CODE_USE_VERTEX=1 is not set.
/rate-limit-optionsOptions to keep working when a claude.ai limit blocks. Hidden; type it in full.
/import [codex|gemini] [--dry-run] [--yes]Pulls in another agent's configuration (OpenAI Codex, Google Gemini CLI): instructions, MCP, commands, subagents, skills. Requires v2.1.213+.

h. MCP prompts as /mcp__server__prompt

An MCP server can expose not only tools and resources, but also prompts that appear as slash commands. Module 11 comes back to this. Remember the form: /mcp__<server>__<prompt> — a github server with a list_prs prompt fires via /mcp__github__list_prs. Since v2.1.199, several skills can chain at the start of a message (up to six) as /skill-a /skill-b do XYZ: every skill is loaded and the final text is passed to them as arguments. This rule applies to skills, not to built-in commands.

Built-in, bundled skill, workflow

A built-in command is wired into the CLI (/status, /config, /permissions, /mcp) and does not appear in /skills. A bundled skill has a prompt as its body (/code-review, /simplify, /verify, /run, /loop, /dataviz): it shows up in /skills, is disabled via skillOverrides, and Claude can invoke it on its own — except /verify (v2.1.215+) and /code-review with the entry "code-review": "user-invocable-only" in settings. A workflow is a bundled script that coordinates several subagents and cross-checks their results (/deep-research); /workflows drives execution, /workflow-authoring loads the reference to write one (v2.1.248+).

Running example: preparing the Kiosque payments PR

Karim has rewritten app/paiements.py to support a second provider. Before opening the PR, we chain a quality ritual that fits in five commands.

# 1. See what will go out
/diff

# 2. Deep local review, with automatic fix for safe items
/code-review high --fix

# 3. Security-specific view
/security-review

# 4. Run the application and verify a payment actually
# works end to end — not just that the tests pass
/verify

# 5. Open the PR
!gh pr create --fill

The review runs as a forked subagent: the main conversation stays usable. Nadia launches in parallel a /subtask "re-read app/notifications.py in light of the new event format" — the result will come back without dragging the whole file read into the conversation.

Later, the team decides to unify the mixed FR/EN naming across app/. The gesture fits in a single call:

/batch rename the functions still in French across app/ (keep English on the API side)
and translate short English comments into French

/batch proposes a breakdown into twelve units, opens a worktree and a PR per unit. Each subagent runs make test before pushing. /tasks tracks the count, claude agents shows the session grid.

Summary

  • /code-review is the center of gravity of review: effort from low to max, --fix, --comment, ultra in the cloud; /review is an exact alias, /security-review covers security.
  • /diff before, /verify after: a review is only worth what it says once confronted with what the program actually does.
  • /subtask, /fork, /background, /batch parallelize across four surfaces: subagent inside the conversation, background copy, full detachment, isolated worktrees.
  • /loop repeats, /goal pursues a condition, /schedule creates a cloud routine, /autofix-pr delegates PR watching to a web session.
  • Extensions (skills, plugins, hooks, MCP) are driven from /skills, /skill-doctor, /reload-skills, /plugin, /reload-plugins, /mcp, /hooks.
  • A bundled skill differs from a built-in command: it shows up in /skills and is disabled via skillOverrides; MCP servers expose prompts as /mcp__server__prompt, and up to six skills can chain at the start of a message since v2.1.199.

Next module: Build your own slash commands: skills from A to Z — build step by step Kiosque's six in-house skills, from frontmatter to shell output injection.