❯ the loop
claude-code codex-cli cursor-cli gemini-cli
> esc

Claude Code / terminal reference

Built from Anthropic's official docs as of 11 July 2026 (Claude Code ~v2.1.207). Click any command to copy it. Things move fast — the source of truth is code.claude.com/docs, and inside a session /help plus typing / always shows what your build actually has.

Cheat sheet Cookbook Changes
what changed · 11 July 2026 · v2.1.207 +1 added ~1 changed all changes →

Recently added

New in the last 30 days. Each row also lives in its own section below.

Elicitation · ElicitationResultnew Around an MCP server's mid-tool-call request for user input; can auto-decide the response. · in Hooks

no matches in this section

Launch

Everything you can type before you're in a session.

claudeStart an interactive session in the current directory.
claude "query"Start interactive with an opening prompt.
claude -p "query"Non-interactive (print mode). Runs the query, prints, exits. The scripting entry point.
cat file | claude -p "query"Pipe stdin through Claude. Pairs with --output-format json.
claude -cContinue the most recent conversation in this directory.
claude -r "<session>" "query"Resume by session ID or name. No args opens the picker.
claude updateUpdate to the latest version.
claude install [version]Install/reinstall the native binary. Accepts stable, latest, or a pinned version.
claude doctorRead-only install + settings diagnostics without starting a session. (In-session /doctor can also fix things.)
claude auth login | logout | statuslogin --console to bill via API instead of subscription. status prints JSON, exit 0 = logged in.
claude setup-tokenGenerate a long-lived OAuth token for CI/scripts.
claude agentsOpen agent view: monitor and dispatch parallel background sessions. --json for scripting.
claude attach <id>Attach to a background session in this terminal.
claude logs <id>Print recent output from a background session.
claude stop <id> / claude rm <id>Stop a background session / remove it from the list (transcript survives).
claude respawn <id>Restart a background session with its conversation intact. --all to pick up a new binary.
claude daemon status | stop --anyInspect/kill the background-session supervisor. --keep-workers leaves sessions running.
claude mcpConfigure MCP servers. claude mcp login <name> runs a server's OAuth flow from the shell.
claude plugin install <name>@<marketplace>Plugin management from the shell (list, enable, disable, update, init, prune).
claude project purge [path]Delete all local state for a project (transcripts, history, debug logs). --dry-run first.
claude ultrareview [target]Deep multi-agent cloud review, non-interactive. Exits 1 on findings — CI-friendly.
claude remote-controlServer mode: drive this machine's Claude Code from claude.ai or the mobile app.
claude auto-mode defaultsPrint the auto-mode classifier's built-in rules as JSON.

no matches in this section

CLI flags

claude --help does not list every flag — absence from help ≠ unavailable.

Session shape

--print, -pNon-interactive mode. Gate for most scripting flags below.
--continue, -c  |  --resume, -rContinue latest / resume by ID or name.
--fork-sessionOn resume, branch to a new session ID instead of writing into the original.
--session-id <uuid>Pin the session ID (useful when you're orchestrating runs).
--name, -n "<name>"Name the session; resume later with claude -r "<name>".
--from-pr <n|url>Resume the sessions linked to a PR.
--worktree, -w [name|#PR]Start inside an isolated git worktree at <repo>/.claude/worktrees/<name>.
--tmuxWith -w: spin the worktree up in a tmux (or iTerm2) pane.
--bg, --backgroundLaunch as a background agent and return immediately. Monitor with claude agents.
--exec '<cmd>'With --bg: run a shell command as a PTY-backed background job (no model).
--add-dir ../pkg-a ../pkg-bExtra working dirs. Grants file access + loads their .claude/skills/, not the rest of their config.
--cloud "task"  |  --teleportPush a task to a web session / pull a web session down into this terminal.

Model & reasoning

--model sonnet|opus|haiku|fable|<id>Session model. Overrides settings and ANTHROPIC_MODEL.
--effort low|medium|high|xhigh|max|ultracodeReasoning effort for this session. ultracode = xhigh + automatic workflow orchestration.
--fallback-model sonnet,haikuOrdered fallback chain when the primary is overloaded or retired.
--advisor opus|sonnet|fableConsult a second model at key moments in a task.
--system-prompt[-file]Replace the entire system prompt. You now own tool guidance and safety text.
--append-system-prompt[-file]Keep the Claude Code prompt, bolt your rules on the end. Usually what you want.
--append-subagent-system-promptSame, but applied to every subagent (print mode only).

Permissions & tools

--permission-mode <mode>default(=manual), acceptEdits, plan, auto, dontAsk, bypassPermissions.
--allowedTools "Bash(git log *)" "Read"Pre-approved rules — the whitelist for autonomous runs.
--disallowedTools "Edit" "mcp__*"Deny rules. A bare tool name removes the tool from context entirely.
--tools "Bash,Edit,Read"Restrict which built-in tools exist at all. "" = none, default = all.
--dangerously-skip-permissions= --permission-mode bypassPermissions. Prefer auto + an allowlist for real work.
--allow-dangerously-skip-permissionsAdds bypass to the Shift+Tab cycle without starting in it.
--permission-prompt-tool <mcp tool>Route permission prompts to an MCP tool in non-interactive runs.

Scripting & output

--output-format text|json|stream-jsonPrint-mode output shape.
--input-format text|stream-jsonFeed structured turns in.
--json-schema '{...}'Validated structured output after the agent finishes (print mode).
--include-partial-messagesToken-level streaming events. Needs -p --output-format stream-json.
--include-hook-eventsEmit every hook lifecycle event into the stream.
--max-turns N  |  --max-budget-usd NHard stops for unattended runs.
--bareSkip discovery of hooks/skills/plugins/MCP/CLAUDE.md. Fastest cold start for scripted calls.
--exclude-dynamic-system-prompt-sectionsMove machine-specific bits out of the system prompt → better cache reuse across machines/users.
--no-session-persistenceDon't write the session to disk (print mode).

Config, extensions, debugging

--settings ./settings.json | '{json}'Override settings keys for this session only.
--setting-sources user,project,localChoose which settings layers load.
--mcp-config ./mcp.json  |  --strict-mcp-configLoad servers from a file / ignore every other MCP config.
--plugin-dir ./my-plugin  |  --plugin-url <zip>Load a plugin for this session only — the fast dev loop for plugin authoring.
--agents '{"reviewer":{...}}'Define subagents inline as JSON (same fields as frontmatter + prompt).
--agent <name>Run the whole session as a named agent.
--disable-slash-commandsKill all skills/commands for the session.
--safe-modeStart with all customization disabled — the bisect tool when your config breaks something.
--debug "api,mcp"  |  --debug-file <path>Category-filtered debug logging. Negate with !statsig.
--verboseFull turn-by-turn output.
--chrome / --no-chrome  |  --ideBrowser integration / auto-connect to a running IDE.
--init | --init-only | --maintenanceRun Setup/SessionStart hooks (with matcher) before — or instead of — a conversation.

no matches in this section

Session & context

Typed inside a session. A command is only recognized at the start of a message.

/clear [name]New conversation, empty context. Old one stays in /resume. Aliases /reset, /new.
/compact [instructions]Summarize the conversation to free context, staying in the same thread.
/context [all]Grid of what's eating the window, with optimization hints. Run this before you blame the model.
/resume [session]Picker or direct resume by ID/name. Background sessions show as bg.
/branch [name]Fork the conversation at this point and switch into the copy. Original preserved.
/rewindRoll code and/or conversation back to a checkpoint, or summarize from a chosen message. Aliases /undo, /checkpoint.
/btw <question>Side question: sees the full conversation, has no tools, never enters history. Press f to fork it into a real session.
/recapOne-line summary of the session so far.
/diffInteractive diff viewer: git diff ↔ per-turn diffs.
/copy [N]Copy the last (or Nth-last) response. Picker for individual code blocks; w writes to a file.
/export [file]Export the conversation as plain text.
/rename [name]Name the session; shows on the prompt bar.
/memoryEdit CLAUDE.md files, toggle and audit auto-memory.
/initGenerate a starter CLAUDE.md. CLAUDE_CODE_NEW_INIT=1 for an interactive flow covering skills + hooks.
/add-dir <path>  |  /cd <path>Grant access to another directory / move the session there (keeps the prompt cache).
/focus  |  /tui [fullscreen]Collapse to prompt + summary + answer / switch the renderer (fullscreen = flicker-free alt-screen).
/exitQuit. In an attached background session this detaches and leaves it running.

no matches in this section

Model, cost & settings

/model [model]Switch model and save as default. Press s on a row for session-only. ←/→ adjusts effort.
/effort [level|auto]low → medium → high → xhigh → max → ultracode. Applies immediately, mid-turn.
/fast [on|off]Fast mode — lower latency, higher token cost.
/advisor [model|off]Second-model consult at key decision points.
/config [key=value]Settings UI, or set directly: /config theme=dark, /config thinking=false. /config --help lists keys.
/statusVersion, model, account, connectivity. Works while Claude is responding.
/usageCost, plan limits, and usage broken down by skill / subagent / plugin / MCP server. Aliases /cost, /stats.
/insightsReport on your own sessions: project areas, interaction patterns, friction points.

no matches in this section

Review & ship

skill = prompt-based bundled skill (Claude can also trigger it on its own). workflow = fans out across subagents in the background.

/code-review [level] [--fix] [--comment] [target]skill Correctness bugs + cleanup review of the diff. Levels low…max, or ultra for the cloud review. --comment posts inline PR comments.
/simplify [target]skill Cleanup-only pass (reuse, simplification, efficiency, abstraction level) — applies fixes, doesn't hunt bugs.
/review [PR]Fast single-pass, read-only review of a GitHub PR. No arg = pick from open PRs.
/security-reviewScan the branch diff for injection, authz, and data-exposure risks.
/ultrareview [PR]Deep multi-agent review in a cloud sandbox. Now preferred as /code-review ultra.
/verifyskill Build + run the app and observe the result, instead of trusting tests and type checks.
/runskill Launch and drive the app to watch a change actually work.
/run-skill-generatorskill Records how to build/launch this project into .claude/skills/run-<name>/ so /run and /verify stop guessing. Run once per repo.
/autofix-pr [prompt]Cloud session that watches the current branch's PR and pushes fixes when CI fails or reviewers comment.
/install-github-appWire up the GitHub app + Actions workflow for automated PR review.

no matches in this section

Parallel & background work

/plan [description]Enter plan mode from the prompt, optionally with the task attached.
/ultraplan <prompt>Draft the plan in a cloud session, review it in the browser, then execute remotely or send it back to the terminal.
/fork <directive>Background subagent that inherits the full conversation, works on the directive, and reports back. (Contrast /branch, which switches you into a copy.)
/batch <instruction>skill Decompose a codebase-wide change into 5–30 units, one background subagent per unit, each in its own worktree, each opening a PR.
/background [prompt]Detach the whole session to keep running headlessly; frees the terminal. Alias /bg. Monitor via claude agents.
/tasksEverything running in the background of this session. Alias /bashes.
/workflowsWatch / pause / resume / save running dynamic workflows.
/goal [condition|clear]Keep working across turns until a condition is met.
/loop [interval] [prompt]skill Re-run a prompt on a cadence while the session is open. /loop 5m check if the deploy finished. Defaults to .claude/loop.md.
/schedule [description]Create routines that run on Anthropic-hosted infra (cron, API, or GitHub triggers). Alias /routines.
/deep-research <question>workflow Fan out web searches, cross-check sources, synthesize a cited report.
/stopStop the attached background session (transcript and worktree kept).

no matches in this section

Config & extensions

/permissionsAllow / ask / deny rules by scope, working dirs, and recent auto-mode denials. Alias /allowed-tools.
/fewer-permission-promptsskill Mines your transcripts for repeated read-only calls and writes an allowlist into project settings.
/hooksView hook configuration for tool events.
/mcp [reconnect|enable|disable <server>]MCP servers + OAuth. In -p mode with no args it prints a status summary.
/plugin [list|install|enable|disable]Plugin menu or direct subcommand.
/reload-plugins [--force]Apply plugin changes without restarting. --force when it would invalidate the prompt cache.
/reload-skillsRe-scan skill + command directories mid-session. (SKILL.md text edits are already hot-reloaded.)
/skillsList skills. Type to filter, t sorts by token cost, Space cycles visibility → writes skillOverrides.
/agentsAs of v2.1.198 this just points you at .claude/agents/ — ask Claude to create/manage subagents, or edit the files.
/doctorskill Setup checkup that can fix: dupe installs, PATH, broken settings, unused skills/MCP/plugins vs their context cost, slow hooks, CLAUDE.md dedupe. Alias /checkup.
/debug [description]skill Turn on debug logging now and analyze the session debug log.
/statusline  |  /theme  |  /keybindingsStatus line (describe what you want), color theme, and your keybindings file.
/sandboxToggle OS-level sandboxing (filesystem + network isolation) where supported.
/terminal-setupInstall Shift+Enter bindings for VS Code, Cursor, Alacritty, Zed.
/dataviz [request]skill Chart/dashboard design guidance with colorblind + contrast validation.
/claude-api [migrate]skill Loads Claude API reference for your language; migrate upgrades existing API code to a newer model.
/design-sync [hint]skill Ship your repo's React design system to Claude Design so generated designs use real components.
/team-onboardingTurn 30 days of your usage into a setup guide a teammate can paste as a first message.
/mcp__<server>__<prompt>MCP servers can expose prompts as commands. Discovered dynamically.

no matches in this section

Platforms & handoff

/desktopContinue this session in the Claude Code desktop app. Alias /app.
/teleportPull a Claude Code web session into this terminal (branch + conversation). Alias /tp.
/remote-controlDrive this local session from claude.ai or the mobile app. Alias /rc.
/web-setup  |  /remote-envConnect GitHub for cloud sessions / pick the default cloud environment.
/ide  |  /chromeIDE integration status / Chrome browser automation settings.
/voice [hold|tap|off]Voice dictation (hold or tap Space).
/install-slack-app  |  /mobileSlack integration / QR code for the mobile app.

no matches in this section

Keyboard & input modes

Input prefixes

/Command or skill menu — built-ins, bundled + your skills, plugin and MCP commands. You can stack up to 6 skills: /code-review /fix-issue 123.
!Shell mode. Runs the command, adds output to context, and (since v2.1.186) Claude responds to it. ! npm test → explanation of the failures.
@File path autocomplete / mention.
TabAccept the greyed-out prompt suggestion (also completes paths in shell mode).

Control

EscInterrupt Claude mid-turn and redirect. Work so far is kept.
Esc EscClear the draft (saved to history) — or, on an empty prompt, open the rewind menu.
Shift+TabCycle permission modes.
Ctrl+CInterrupt; on an idle prompt, first press clears input, second exits.
Ctrl+DExit the session.
Ctrl+OToggle the transcript viewer (expands tool calls and collapsed MCP output).
Ctrl+TToggle the task list.
Ctrl+BBackground the running bash command or agent. (Twice under tmux.)
Ctrl+X Ctrl+KStop all background subagents in this session (press twice to confirm).
Ctrl+RReverse-search prompt history. Ctrl+S cycles scope: session → project → all.
Ctrl+G / Ctrl+X Ctrl+EOpen the prompt in $EDITOR.
Ctrl+VPaste an image from the clipboard as an [Image #N] chip. (Alt+V on Windows/WSL.)
Ctrl+LRedraw the screen when the terminal gets garbled.
Alt/Option+PSwitch model without losing your prompt.
Alt/Option+TToggle extended thinking. (No effect on Fable 5 — always thinks.)
Alt/Option+OToggle fast mode.
Shift+Enter / \ + Enter / Ctrl+JNewline. Shift+Enter is native in iTerm2, Ghostty, WezTerm, Kitty, Warp, Terminal.app, Windows Terminal.

no matches in this section

Permission modes

Cycle with Shift+Tab, set at launch with --permission-mode, or persist as defaultMode in settings.

default (aka manual)Ask before each new tool use.
planRead-only analysis; produces a plan you approve before anything is written.
acceptEditsFile edits auto-approved; other tools still prompt.
autoA classifier approves safe actions and blocks risky ones. The sane default for long autonomous runs — tune it with claude auto-mode defaults and review denials in /permissions.
dontAskOnly pre-approved tools run; everything else is denied rather than prompted.
bypassPermissionsNo checks at all. Reach for a sandbox or container before you reach for this.

no matches in this section

Skills authoring

Custom commands merged into skills: .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both produce /deploy. Skills add supporting files, invocation control, and automatic loading. Follows the agentskills.io standard.

Where they live (precedence: enterprise → personal → project; a same-named skill overrides a bundled one)

~/.claude/skills/<name>/SKILL.mdPersonal — all your projects.
.claude/skills/<name>/SKILL.mdProject — commit it. Also loads from parent dirs up to the repo root, and from nested package dirs on demand (apps/web:deploy).
<plugin>/skills/<name>/SKILL.mdPlugin — namespaced /plugin-name:skill-name, so it can't collide.
.claude/commands/<name>.mdLegacy custom command. Still works, same frontmatter. Skill wins on a name clash.

Frontmatter

descriptionThe only field that really matters — Claude matches on it. Key use case first; combined with when_to_use it's truncated at 1,536 chars.
when_to_useExtra trigger phrases / example requests, appended to the description.
disable-model-invocation: trueOnly you can fire it. Use for anything with side effects (/deploy, /commit). Also keeps its description out of context.
user-invocable: falseOnly Claude can fire it. Use for background knowledge that isn't a meaningful command.
allowed-tools: Bash(git add *) ReadPre-approves those tools while the skill is active. Doesn't restrict anything else.
disallowed-toolsRemoves tools from the pool while active (e.g. AskUserQuestion for an unattended loop). Clears on your next message.
context: fork  +  agent: ExploreRun the skill as a subagent — SKILL.md becomes the prompt. Only makes sense for skills with a real task in them.
model / effortOverride for the duration of the turn. Session settings resume on your next prompt.
paths: "ios/**"Only auto-activate when working on matching files.
argument-hint / argumentsAutocomplete hint, and named positional args for $name substitution.
hooksHooks scoped to this skill's lifecycle.

Substitutions & dynamic context

$ARGUMENTS  /  $0 $1  /  $nameAll args / positional / named. If $ARGUMENTS is absent, args are appended as ARGUMENTS: ….
${CLAUDE_SKILL_DIR}The skill's own directory — use it to call bundled scripts regardless of cwd.
${CLAUDE_PROJECT_DIR}Repo root. Works in the body and in allowed-tools rules.
${CLAUDE_SESSION_ID} / ${CLAUDE_EFFORT}Session ID for logging; current effort level for adapting instructions.
!`git diff HEAD`Runs before Claude sees the skill; output is inlined. Multi-line: a fenced ```! block. Only recognized at line start or after whitespace.
ultrathinkPut the word anywhere in the skill body to request deeper reasoning for that run.

Minimal SKILL.md

---
name: summarize-changes
description: Summarize uncommitted changes and flag risks. Use when the user asks
  what changed, wants a commit message, or asks to review their diff.
allowed-tools: Bash(git *)
---

## Current changes
!`git diff HEAD`

## Instructions
Summarize in 2–3 bullets, then list risks: missing error handling,
hardcoded values, tests that need updating.

Gotchas worth internalizing: skill content enters context once and stays for the session (write standing instructions, not one-time steps) — keep SKILL.md under ~500 lines and push reference material into sibling files. If you have many skills, descriptions get truncated to a context budget (1% of the window; raise with skillListingBudgetFraction); /doctor tells you who's eating it. Turn skills off per-project without editing them via skillOverrides in .claude/settings.local.json.

no matches in this section

Subagents & teams

.claude/agents/<name>.mdSubagent definition: frontmatter (description, tools, model, permission mode, skills to preload, hooks) + a body that becomes its system prompt.
Explore / Plan / general-purposeBuilt-in agents. Explore and Plan skip CLAUDE.md and git status to stay small — ideal for parallel research.
Subagent vs /btwSubagent = full tools, empty context (go find out something new). /btw = full context, no tools (ask about what's already known).
Agent teamsMultiple teammates with shared task claiming, plan approval, and quality-gate hooks. Display mode via --teammate-mode in-process|auto|tmux|iterm2.
IsolationSubagents can each get their own git worktree; /batch does this automatically per unit of work.
CostSubagents don't share the parent's prompt cache. Use them to isolate high-volume reads, not to generate parallel opinions.

no matches in this section

Hooks

Shell, HTTP, prompt-based, or agent-based handlers on lifecycle events. Configure in settings under hooks, in a plugin, or in skill/agent frontmatter. Exit code 2 blocks; JSON output can inject context or make decisions. Inspect with /hooks, debug with claude --debug hooks.

SessionStart · Setup · SessionEndSession lifecycle. Setup supports init / maintenance matchers via CLI flags.
UserPromptSubmit · UserPromptExpansionIntercept or enrich the prompt before it reaches the model.
PreToolUse · PostToolUse · PostToolUseFailure · PostToolBatchAround tool calls. Filter with matchers and the if field (tool name + args). Auto-format on edit lives here.
PermissionRequest · PermissionDeniedAuto-approve or auto-deny specific permission prompts.
Stop · StopFailure · Notification · MessageDisplayTurn boundaries and notifications (terminal bell, Slack ping, etc.).
SubagentStart · SubagentStop · TeammateIdleAgent lifecycle — quality gates for teams.
TaskCreated · TaskCompletedTask-list events.
PreCompact · PostCompact · InstructionsLoadedRe-inject context after compaction; audit what loaded.
FileChanged · CwdChanged · ConfigChangeReact to the environment moving under you.
WorktreeCreate · WorktreeRemoveCopy gitignored files (.env, xcconfig, local secrets) into new worktrees.
Elicitation · ElicitationResultnew Around an MCP server's mid-tool-call request for user input; can auto-decide the response.

no matches in this section

File map

~/.claude/settings.jsonUser settings: model, hooks, permissions, statusline.
.claude/settings.jsonProject settings — commit these.
.claude/settings.local.jsonPersonal project overrides — gitignored. skillOverrides lands here.
CLAUDE.md / AGENTS.mdAlways-loaded project memory. Layer per-directory in large repos; move procedures out to skills.
.claude/rules/Path-scoped rules — the alternative to one bloated CLAUDE.md.
.claude/skills/  ·  .claude/commands/Skills (dir + SKILL.md) and legacy commands (single .md).
.claude/agents/Subagent definitions.
.mcp.jsonProject-scoped MCP servers (shared with the team).
.claude/worktrees/Where -w and /batch put isolated worktrees.
.claude/loop.mdDefault prompt for /loop.
~/.claude/projects/<slug>/Session transcripts as JSONL — tail these to watch an autonomous run.
~/.claude/themes/  ·  keybindingsCustom themes; /keybindings opens your bindings file.

no matches in this section