GeezerVison

Claude Code is Anthropic’s AI-powered CLI tool that reads your codebase, edits files, runs commands, and manages git workflows directly from the terminal. It operates as an agentic coding assistant: you describe what you want in plain English, and it figures out which files to read, what commands to run, and what changes to make. This cheat sheet covers every command, shortcut, configuration option, and workflow pattern you need to be productive with Claude Code.

Original content from computingforgeeks.com - post 162997

The releases since 2.1.282 moved the aliases again: sonnet now resolves to Sonnet 5.5 and haiku to Haiku 5.5, while opus stays on Opus 5.5, and 2.1.321 moved fable from Fable 5.1 to Fable 5.2, which is cheaper at $8/$40 per million tokens and supports /fast. Every plan and provider now starts interactive sessions in auto mode when no permission mode is configured, Ultracode became its own /effort ultracode switch, /doctor prompt-audit flags CLAUDE.md rules written for older models, and claude project purge is now plain claude purge. Every alias, command, and flag below was checked against Claude Code 2.1.303 on macOS in October 2026, the cost and error output comes from that install, and the 2.1.304 compact-window change was checked on an Ubuntu 26.04 VM later that month. The fable alias, the /model picker and the Fable fast-mode dialog were rechecked on Claude Code 2.1.322 on an Ubuntu 24.04 VM in November 2026.

This cheat sheet is also available as a GitHub repository you can star, fork, and reference offline. Contributions and suggestions are welcome there. For the official changelog and new features, check the Claude Code releases on GitHub.

Claude Code Commands Cheat Sheet: the Daily 20

If you only learn twenty things, learn these. They cover starting and resuming sessions, steering the model, keeping the context window clean, and getting out of trouble, which is most of what a working day in Claude Code involves:

Action	How
Start new session	claude
Continue last session	claude -c
Initialize project	/init
Switch to plan mode	Shift+Tab
Switch model	/model opus or Option+P / Alt+P
Undo changes	Esc + Esc
Free context space	/compact
Start fresh	/clear
Check plan usage and cost	/usage (/cost is an alias)
Review changes	/diff
Security scan	/security-review
Name session	/rename my-feature
Resume session	/resume or claude -r
Mention a file	@path/to/file
Run bash directly	!ls -la
Multiline input	\ + Enter
Stop Claude mid-turn	Esc (Ctrl+C also works, but twice exits)
Exit	Ctrl+D
Non-interactive	claude -p "query"
Show all commands	/help

Fix-It Recipes: What to Press When Something Goes Wrong

Most cheat sheets list commands alphabetically, which is no help when Claude has just rewritten the wrong file. This table starts from the situation instead. Every key and command in it was checked against 2.1.303:

Situation	Do this
Claude is editing the wrong thing right now	Esc stops the turn and keeps what was done so far
Undo the last edits, code and chat together	Esc Esc on an empty prompt, or /rewind, then pick a checkpoint
You cleared a long prompt with Ctrl+C by accident	Press Up while the prompt is still empty to bring the draft back
Park a half-written prompt to ask something else	Ctrl+S stashes it; press it again on an empty prompt to restore
Quick question without polluting the context	/btw what does this regex match?
Context is filling up mid-task	/compact keep the auth refactor notes, or /context first to see what is eating it
Try a different approach without losing this one	/branch copies the conversation so you can go back
Switching to unrelated work	/clear, which is cheaper than compacting
Send queued messages without waiting for the turn	Ctrl+Enter, or Ctrl+X Ctrl+S in terminals without extended keys
Background subagents are running away	Ctrl+X Ctrl+K twice within 3 seconds stops them all
Hit a usage limit	/rate-limit-options lists waiting, credits, and other ways to keep going
Something is broken in your setup	/doctor, then f to let Claude fix what it finds; claude --safe-mode starts with every customization off
CLAUDE.md rules feel ignored on newer models	/doctor prompt-audit flags instructions written for older models
Find the session you named yesterday	claude -r auth-fix resumes an exact name match directly; a partial name opens the picker pre-filtered

The two people get wrong most often are Esc and Ctrl+C. Esc interrupts Claude and leaves you in the session, while a second Ctrl+C on an empty prompt exits Claude Code entirely. If you came from a shell where Ctrl+C is the reflex for “stop”, retrain that finger, because exiting throws away the in-memory state of any background work the session was supervising.

What Changed From 2.1.282 to 2.1.304

These are the changes since the last revision of this cheat sheet that alter how you type or what you get by default. The full list lives in the official changelog, but most of it is bug fixes:

Change	Version	What it means for you
Auto mode is the starting mode on every plan and provider when no mode is configured	2.1.283, 2.1.284	Pro, Max, and Team already had it; Enterprise, API, and third-party provider users now get it too. Set permissions.defaultMode for the old prompting behavior
/auto-mode-setup works on every provider	2.1.304	The guided setup that drafts autoMode.environment from your project now also runs on Bedrock, Vertex, Foundry, the Claude apps gateway, and sessions with telemetry turned off
claude -p and Agent SDK runs with no mode set start in manual (default) mode	2.1.302	Tool calls that need approval are denied in scripts unless you pass --allowedTools or --permission-mode
/doctor prompt-audit (also /checkup prompt-audit)	2.1.283	Audits CLAUDE.md, skills, agents, and commands for prompting patterns aimed at older models
Ultracode is its own toggle: /effort ultracode on|off	2.1.284	The slash command no longer forces xhigh and stays on at any effort level; the --effort ultracode launch flag still starts at xhigh
/mcp reconnect all	2.1.284	Retries every MCP server that failed to connect or needs authentication
claude --desktop, claude plugin configure	2.1.285	Hand a session to the desktop app; inspect a plugin’s options
Background commands get a time limit in unattended runs only	2.1.285, 2.1.288	In -p, SDK, CI, and cloud runs, background commands stop after 30 minutes by default (raise it with BASH_DEFAULT_TIMEOUT_MS and BASH_MAX_TIMEOUT_MS). Terminal, desktop, and VS Code sessions have no limit since 2.1.288
/code-review --max-findings <n>|all|default	2.1.288	Ask for more or fewer findings; the choice sticks until default
/autocompact saved per model	2.1.288	Opus and Sonnet can keep different compaction windows
CLAUDE_CODE_AUTO_COMPACT_WINDOW accepts k and M suffixes	2.1.304	Before this, 200k silently gave you a 100K window. Values outside 100k to 1M are now clamped with a warning, and unreadable ones are ignored with a warning
claude attach and claude logs accept part of a session name	2.1.290	No need to copy the short id
Idle compaction for conversations over 200K tokens	2.1.299	Terminal sessions idle about 54 minutes get compacted before the prompt cache expires. Change the threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS; turn it off with the /config row or "idleCompaction": false
claude project purge renamed to claude purge	2.1.288	The old name still works and prints a notice. --dry-run lists what would go, --all covers every project
MCP tool search on by default behind a custom ANTHROPIC_BASE_URL	2.1.304	Gateway users now get deferred MCP tools without setting ENABLE_TOOL_SEARCH. If your gateway cannot forward tool_reference blocks, set ENABLE_TOOL_SEARCH=false. On 2.1.304 itself a gateway that rejects tool search fails every request; 2.1.305 and later retry without it

The auto-mode default is the one that changes behavior without you noticing. On a fresh install you will see far fewer permission prompts than older tutorials describe, and that is the classifier at work rather than a broken prompt. If your team policy requires a human to approve each command, pin "defaultMode": "default" in the project’s .claude/settings.json so a new laptop does not quietly start in auto. That only sets the starting mode, though, because anyone can still press Shift+Tab into auto; real enforcement is permissions.disableAutoMode set to "disable" in managed settings.

The compact window fix is worth checking if you set the variable in a shell profile months ago. Before 2.1.304, CLAUDE_CODE_AUTO_COMPACT_WINDOW=200k silently set a 100K window, so sessions compacted at half the size you asked for and nothing said why. From 2.1.304 a bad value tells you what it did instead of failing silently. This is the 2.1.304 binary on an Ubuntu 26.04 VM with a value below the floor:

CLAUDE_CODE_AUTO_COMPACT_WINDOW=50k claude -p "say hi"

The warning names the value it read and the size it used, before the session continues (that VM was not signed in, hence the last line):

CLAUDE_CODE_AUTO_COMPACT_WINDOW="50k" reads as 50,000 tokens, so Claude Code treats it as the 100k minimum. Set it to a size from 100k to 1M, such as 200k or 200000, or unset it.
Not logged in · Please run /login

2M gets the matching “is too large, so Claude Code treats it as the 1M maximum” line, and a typo such as abc is ignored with a warning rather than guessed at. 200k, 500K, 1M, and the plain 200000 all pass without a word. The same k and M forms work in the /autocompact command and the --autocompact flag.

Installation

Install Claude Code on macOS, Linux, or Windows (WSL):

curl -fsSL https://claude.ai/install.sh | bash

On macOS via Homebrew:

brew install --cask claude-code

On Windows (PowerShell):

irm https://claude.ai/install.ps1 | iex

Verify and authenticate:

claude --version
claude auth login

A working install prints the version followed by (Claude Code), for example 2.1.304 (Claude Code) on the Ubuntu lab VM. Sign-in is a one-time step per machine; later sessions reuse it.

Keyboard Shortcuts

These work in the interactive Claude Code terminal. This is the most-referenced section, so bookmark it.

Essential Shortcuts

Shortcut	Action
Ctrl+C	Interrupt a running operation. If nothing is running, first press clears input, second exits. Up on the empty prompt brings a cleared draft back
Ctrl+S	Stash the current prompt; press again on an empty prompt to restore it with its mode and pasted content
Ctrl+Enter	Send queued messages now instead of after the current turn (Ctrl+X Ctrl+S in any terminal)
Ctrl+_	Undo the last edit in the prompt input
Esc	Interrupt Claude mid-turn, keeping the work done so far
Esc + Esc	Clear the input draft, or (when input is empty) open the rewind menu to restore a checkpoint
Ctrl+D	Exit Claude Code
Ctrl+L	Redraw the screen to recover a garbled display
Ctrl+R	Reverse search command history
Shift+Tab	Cycle permission modes (manual, acceptEdits, plan, plus auto and bypassPermissions when available). New sessions with no configured mode start in auto
Ctrl+G	Open input in external editor (also Ctrl+X Ctrl+E)
Ctrl+O	Toggle the transcript viewer (detailed tool usage, expands collapsed MCP calls)
Ctrl+T	Toggle the task list (on Opus 4.8, Sonnet 5, and newer the task tools need CLAUDE_CODE_ENABLE_TODO_TOOLS=1)
Ctrl+B	Move a running bash command or agent to the background (tmux: press twice)
Ctrl+X Ctrl+K	Kill all running background subagents (press twice within 3 seconds)

Input and Editing

Shortcut	Action
\ + Enter	New line, works in every terminal
Ctrl+J	New line, works in any terminal with no configuration
Option+Enter (Mac) / Alt+Enter (Linux/Win)	New line (platform default)
Shift+Enter	New line in iTerm2, WezTerm, Ghostty, Kitty, Warp, Apple Terminal, Windows Terminal. Run /terminal-setup for VS Code, Cursor, Zed
Ctrl+A / Ctrl+E	Move cursor to start / end of line
Ctrl+K	Delete to end of line
Ctrl+U	Delete from cursor to line start
Ctrl+W	Delete back to the previous whitespace, Bash style: one press removes a whole path or --flag=value. Option+Delete (macOS) or Ctrl+Backspace (Windows) deletes a single word
Ctrl+Y	Paste text deleted with Ctrl+K/U/W
Alt+B / Alt+F	Move cursor back/forward one word (macOS needs Option set as Meta)
Cmd+V (Mac) / Ctrl+V or Alt+V (Linux/Win/WSL)	Paste image from clipboard
Up/Down	Move cursor, then navigate command history at the edges

Model and Mode Switching

Shortcut	Action
Option+P (Mac) / Alt+P (Linux/Win)	Open the model picker (Opus, Sonnet, Fable, Haiku)
Option+T (Mac) / Alt+T (Linux/Win)	Toggle extended thinking (no effect on Opus 5.5, Sonnet 5.5, Haiku 5.5, or the Fable models, which always think)
Option+O (Mac) / Alt+O (Linux/Win)	Toggle fast mode
Shift+Tab / Alt+M	Cycle permission mode (all platforms)
?	Show all shortcuts (in the transcript viewer)

Quick Prefixes

Prefix	Action
/	Open slash command menu
!	Run bash command directly
@	Autocomplete file path mention

Slash Commands Reference

Session and Context Management

Command	What It Does
/clear	Wipe conversation history and free context window
/compact [focus]	Summarize conversation to free context, optionally focused on one topic
/cost	Alias for /usage: plan limits, session cost, and prompt-cache hits
/context	Visualize what is consuming your context window
/resume	Pick and resume a previous session
/rename [name]	Name current session for easy resuming later
/rewind	Rewind to a previous checkpoint
/diff	Review uncommitted changes in the working tree. In fullscreen mode it opens a panel beside the conversation that updates as Claude edits
/copy	Copy last response to clipboard
/export [file]	Export conversation as text file
/cd [path]	Move this session to a new working directory, keeping the prompt cache
/background / /bg	Detach the whole session to keep running as a background agent
/tasks	View and manage everything running in the background, including finished subagents and the model each ran on (alias /bashes)
/autocompact [size|auto]	Set how full the context gets before auto-compaction, for example /autocompact 500k

Configuration and Model

Command	What It Does
/model [name]	Switch model (opus, sonnet, haiku, fable) and save it as the default for new sessions. Press s in the picker to use it for this session only
/effort [level]	Set reasoning effort (low, medium, high, xhigh, max, ultracode, auto; status prints it). Saved per model, so each model keeps its own level. Opus 5.5, Sonnet 5.5, and Haiku 5.5 default to medium, Opus 4.7 to xhigh, the rest to high
/config [key=value]	Open settings, or set a key inline (/config theme=dark, /config model=sonnet). Vim editor mode now lives here. Run /config --help for keys
/permissions	View and manage tool permissions and review recent auto-mode denials
/mcp	Manage MCP server connections and OAuth
/terminal-setup	Configure terminal keybindings for your shell
/keybindings	Open your keyboard shortcuts file to rebind keys
/theme	Change color theme (light, dark, auto, colorblind-safe, custom)
/tui [fullscreen]	Switch the terminal renderer; fullscreen enables the flicker-free view
/output-style [name]	List output styles or switch to one, for example /output-style concise
/skills	List skills with their token cost (t sorts by tokens) and toggle each one’s visibility
/plugin / /reload-plugins	Open the plugin menu, or reload active plugins without restarting

Project and Code Tools

Command	What It Does
/init	Initialize project, generating a CLAUDE.md with project context
/memory	View and edit CLAUDE.md files and auto-memory
/code-review [level] [--fix] [--comment]	Review the diff for bugs and cleanups. --fix applies findings, --comment posts inline PR comments, ultra runs a cloud multi-agent review
/simplify [target]	Cleanup-only review (reuse, simplification, efficiency) that auto-applies fixes. Use /code-review to hunt bugs
/review [PR]	Review a pull request locally in the current session
/security-review	Scan pending changes for security vulnerabilities
/run	Launch and drive your app to confirm a change actually works, not just that tests pass
/verify	Build, run, and observe the app to confirm a change does what it should. Claude no longer runs this (or /code-review) on its own; invoke it explicitly
/advisor [model|off]	Consult a second model (fable, opus, sonnet, or a model ID) at key moments during a task
/skill-doctor	Show which loaded skills go unused and what each costs in context, so you can prune them. Needs feature-flag fetching enabled
/import [codex|gemini|cursor]	Bring instruction files, MCP servers, commands, subagents, and skills over from another coding agent (--dry-run previews). Not available on Bedrock, Vertex, or Foundry
/add-dir [path]	Add another directory to the working context
/plan	Enter plan mode, a read-only exploration before making changes
/branch [name]	Branch the current conversation to try a different direction without losing this one
/fork [directive]	Copy the conversation into a new background session (its own row in claude agents) while you keep working in the original
/subtask [directive]	Spawn an in-session subagent that inherits the full conversation and reports back when done (the previous /fork behavior)
/goal [condition]	Keep working across turns until a stated condition is met
/batch [task]	Decompose a large change into 5 to 30 units and run each in its own parallel worktree
/debug [description]	Troubleshoot problems in the current session

System and Account

Command	What It Does
/help	Show all available commands
/doctor / /checkup	Full setup checkup: diagnoses install, settings, auth, and MCP, then press f to have Claude fix what it finds
/insights	Generate a session analysis report: patterns, efficiency, suggestions
/status	Show version, model, and account info
/login / /logout	Sign in or out
/bug	Submit a bug report with session logs
/stats	Open /usage on its Stats tab: usage patterns, daily streaks, and session history
/usage	Show plan usage and rate limits
/release-notes	View changelog for current version
/loop [interval] [prompt]	Run a prompt repeatedly while the session is open. Omit the interval and Claude paces itself
/voice	Toggle push-to-talk voice dictation (hold Space)
/schedule	Create, update, list, or run scheduled remote agents (cron tasks)
/fast	Toggle fast mode (same model, faster output, higher price) on Opus 5.5 or Fable 5.2. On Fable, run it while Fable 5.2 is the active model
/hooks	View and manage configured hooks
/desktop	Hand off current session to the Desktop app
/teleport	Pull a web/iOS session into your terminal
/web-setup	Connect GitHub for cloud sessions using your local gh credentials
/list-agents	List the subagents, teammates, and other sessions Claude can message (alias /peers). Only where cross-session messaging is on
/remote-control	Continue this local session from claude.ai or the Claude app on another device
/reload-skills	Re-scan skill and command directories so on-disk changes apply without a restart
/sandbox	Toggle sandbox mode (supported platforms only)
/recap	Summarize what has happened in the session so far
/statusline	Configure the status line from a description or your shell prompt
/privacy-settings	View and update privacy settings (Pro and Max)
/focus	Toggle the focus view: only your last prompt, a one-line tool summary, and the final response
/powerup	Learn Claude Code features through quick interactive lessons with animated demos
/dataviz	Design charts and dashboards with colorblind-safe, contrast-checked palettes
/deep-research [question]	Fan out web searches, cross-check sources, and synthesize a cited report
/autofix-pr [prompt]	Spawn a cloud session that watches your PR and pushes fixes when CI fails or reviewers comment
/fewer-permission-prompts	Scan your transcripts and pre-approve common read-only commands to cut prompt fatigue
/team-onboarding	Generate a team onboarding guide from your last 30 days of Claude Code usage
/rate-limit-options	Ways to keep working when a usage limit blocks a request: wait and auto-continue, add usage credits, or upgrade your plan
/usage-credits	Configure usage credits, or request them from your admin on Team and Enterprise
/auto-mode-setup	Draft autoMode.environment entries from your project and recent sessions, review, then save
/ultrareview [PR or branch]	Deep multi-agent code review in a cloud sandbox
/stop	Stop the background session you are attached to; transcript and worktree are kept
/exit / /quit	Exit the CLI. In an attached background session it detaches and the session keeps running
/color [color]	Color the prompt bar for this session (red, blue, green, yellow, purple, orange, pink, cyan)
/ide / /chrome	Manage IDE integrations, or Claude in Chrome settings
/install-github-app	Install the Claude GitHub App on a repo, optionally with GitHub Actions workflows
/remote-env	Choose the default cloud environment for cloud sessions started from the CLI
/setup-bedrock / /setup-vertex	Provider wizards, hidden until CLAUDE_CODE_USE_BEDROCK=1 or the Vertex equivalent is set
/update-config [request]	Bundled skill: describe a settings change and Claude edits the right settings.json
/vim, /pr-comments	Removed (2.1.92 and 2.1.91). Vim mode lives in /config under Editor mode; ask Claude directly for PR comments

CLI Flags – Non-Interactive and Automation

Starting Sessions

Claude Code reads your current directory as the project root. Launch it from the repo where you want work to happen:

# Start interactive session
claude

# Start with initial prompt
claude "refactor the auth module"

# Continue most recent session
claude -c

# Resume specific session by name
claude -r "auth-refactor"

# Name a session for later
claude -n "feature-payments"

# Run in isolated git worktree
claude -w feature-branch

# Run in worktree with tmux panes
claude -w feature-branch --tmux

# Resume session linked to a GitHub PR
claude --from-pr 42

# Fork a session (new ID, keeps context)
claude -c --fork-session

# Enable auto mode (AI decides permissions)
claude --permission-mode auto

# Skip all permission prompts (sandboxes and containers only)
claude --dangerously-skip-permissions

# Add bypassPermissions to the Shift+Tab cycle without starting in it
claude --permission-mode plan --allow-dangerously-skip-permissions

# Bare mode (skip hooks, plugins, auto-memory) for fast scripted calls
claude --bare

# Safe mode: start with ALL customizations disabled to troubleshoot a broken config
claude --safe-mode

# Restricted mode: no Bash, code-running tools, or WebFetch; file tools confined to the working dirs;
# only managed settings and --settings load, so your own hooks and permission rules are skipped
claude --restricted

# Continue this local session from claude.ai or the Claude app on another device
claude --remote-control

# Preview, then import config from another coding agent (codex, gemini, or cursor)
claude import codex --dry-run

Run the --dry-run form of claude import first on any repo you care about, since it only lists what would be brought over.

Print Mode (Non-Interactive / Piping)

The -p flag runs Claude Code non-interactively: it processes the prompt and exits. This is the key to integrating Claude Code into scripts, CI/CD pipelines, and Unix pipes:

# Simple query
claude -p "explain this error in server.log"

# Pipe input
cat error.log | claude -p "what caused this crash?"

# Git diff review
git diff main | claude -p "review for security issues"

# JSON output for scripting
claude -p "list all TODO comments" --output-format json

# Structured output with schema validation
claude -p "extract function names from auth.py" \
  --output-format json \
  --json-schema '{"type":"object","properties":{"functions":{"type":"array","items":{"type":"string"}}}}'

# Auto-approve tools (careful with this)
claude -p "run tests and fix failures" --allowedTools "Bash,Read,Edit"

# Set spending limit
claude -p "refactor the API layer" --max-budget-usd 5.00

# Limit turns
claude -p "fix the bug" --max-turns 3

# Fallback model when primary is overloaded
claude -p "review this code" --fallback-model haiku

# Restrict available tools
claude -p "analyze this" --tools "Read,Grep,Glob"

# Block specific tools (deny list)
claude -p "refactor auth module" --disallowedTools "Bash(rm:*),Bash(sudo:*)"

# Override system prompt entirely
claude -p "review this code" --system-prompt "You are a security auditor. Focus only on vulnerabilities."

# Append to default system prompt
claude -p "explain this function" --append-system-prompt "Always include time complexity analysis."

# Load MCP servers from external config file
claude -p "check open issues" --mcp-config ./ci-mcp.json

# Disable session persistence (ephemeral, for CI/CD)
claude -p "run lint check" --no-session-persistence

# Debug mode (logs API calls, tool usage, timings)
claude -p "fix the build" --debug

# Debug with category filter
claude -p "fix the build" --debug "api,hooks"

# Write debug output to a file
claude -p "fix the build" --debug-file /tmp/claude-debug.log

# Select a specific agent
claude -p "review auth.py" --agent code-reviewer

# Streaming JSON output (realtime events for SDK and automation)
claude -p "fix the build" --output-format stream-json --verbose

# Streaming JSON input
claude -p --input-format stream-json --output-format stream-json

# Pin a specific session ID (must be a valid UUID)
claude -p "run lint" --session-id 550e8400-e29b-41d4-a716-446655440000

# Load extra settings inline or from a file (overrides matching keys for this run)
claude -p "audit deps" --settings ./ci-settings.json

# Only load specific setting sources
claude -p "check types" --setting-sources user,project

# Replace the default system prompt from a file
claude -p "review" --system-prompt-file ./reviewer.txt

# Improve prompt-cache reuse across users and machines in scripted runs
claude -p "lint" --exclude-dynamic-system-prompt-sections

# Enable Chrome browser integration
claude --chrome

Every one of those runs bills the system prompt, tool definitions, and your CLAUDE.md files before it reads your prompt, so the model alias matters more than the prompt length for short scripted calls.

What One claude -p Call Costs per Model

To put numbers on that, the same one-word prompt went to each alias from a scratch directory, with JSON output so the cost and the resolved model ID come back with the answer:

claude -p --model haiku --output-format json "Reply with the single word ok" | jq '{result, total_cost_usd, duration_ms, num_turns, modelUsage: (.modelUsage | keys)}'

The modelUsage key shows which model the alias resolved to, which is the quickest way to confirm what haiku means on your account today:

{
  "result": "ok",
  "total_cost_usd": 0.00342609,
  "duration_ms": 1173,
  "num_turns": 1,
  "modelUsage": [
    "claude-haiku-5-5"
  ]
}

Repeating it for every alias on the same setup gave this spread. The first call in a directory also pays to write the prompt cache, which is why a repeat Haiku call came in slightly cheaper than the first:

Alias	Resolved model ID	Cost of a one-word reply
haiku	claude-haiku-5-5	$0.0034 to $0.0039
sonnet	claude-sonnet-5-5	$0.074
opus	claude-opus-5-5	$0.143
fable (on 2.1.303)	claude-fable-5-1	$0.404

Those figures come from 2.1.303, when fable still meant Fable 5.1 at $10/$50 per million tokens. From 2.1.321 the same alias calls claude-fable-5-2, which the /model picker lists at $8/$40, so expect the Fable row to come in lower on a current install. Run the jq command above with --model fable to get your own figure. On 2.1.303, Fable cost about 100 times what Haiku did for the same trivial answer. A cron job that runs claude -p every five minutes is 288 calls a day, which works out to about $41 a day on the default Opus 5.5 and about $1 on Haiku before it does any real work, so scripted checks should name --model haiku explicitly and set --max-budget-usd as a ceiling. Save Opus and Fable for runs where the reasoning is the point.

Error: “Not logged in · Please run /login” with –bare

--bare skips hooks, plugins, and CLAUDE.md discovery, which makes CI runs reproducible, but it also stops reading OAuth tokens and the keychain. On a machine signed in with a Claude subscription, the same command that works normally fails immediately:

claude -p --bare --model haiku "Reply ok"

The session never starts:

Not logged in · Please run /login

Running /login will not help. For the Anthropic API, bare mode accepts only ANTHROPIC_API_KEY or an apiKeyHelper passed through --settings (Bedrock, Vertex, and Foundry still read their own provider credentials), so export an API key in the CI environment, or drop --bare and use --setting-sources to control which settings load.

A mistyped subcommand becomes a prompt

Claude Code treats any unknown first word as the prompt. Interactively you notice at once, but when stdout is a pipe or a file the session runs in print mode, so claude stauts | head sends the word “stauts” to the model, bills a full turn, and prints a chatty answer instead of an error. Scripts that wrap Claude Code subcommands should check claude --help output for the subcommand first, or quote the intent explicitly with -p so a typo is at least obvious in the logs.

Model and Effort Flags

Pick a model and reasoning budget up front instead of toggling mid-session. Flags stick for the session they start:

# Use specific model
claude --model opus
claude --model sonnet
claude --model haiku
claude --model fable

# Pin an exact version instead of an alias
claude --model claude-opus-5-5

# Set reasoning effort for this session only
claude --effort high    # More thorough (costs more)
claude --effort low     # Faster, cheaper
claude --effort ultracode   # xhigh plus dynamic workflows

# Custom system prompt
claude --append-system-prompt "Always use TypeScript. Never use any."

# Load system prompt from file
claude --append-system-prompt-file ./coding-rules.txt

Both append flags add to the default system prompt instead of replacing it, so Claude Code keeps its built-in tool instructions. --system-prompt (or --system-prompt-file) is the one that replaces everything.

CLAUDE.md – Project Instructions

CLAUDE.md is the single most important file for customizing Claude Code behavior on a per-project basis. It is loaded automatically at the start of every session. For the full directory structure, including rules, agents, skills, and settings, see the .claude directory guide.

File Locations and Scope

Location	Scope	Shared with Team?
./CLAUDE.md	This project only	Yes (commit to git)
./.claude/CLAUDE.md	This project only	Yes (commit to git)
~/.claude/CLAUDE.md	All your projects	No (personal)
.claude/rules/*.md	Path-specific rules	Yes (commit to git)
./AGENTS.md	Read instead of CLAUDE.md when the repo has no CLAUDE.md	Yes (commit to git)

Repos set up for Codex or other agents often carry an AGENTS.md already. Claude Code reads it on its own when no CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md exists in the working directory or above it. If both files exist, only CLAUDE.md loads, so either import @AGENTS.md from CLAUDE.md or change the Project instructions row in /config to read both.

Example CLAUDE.md

Create the file at the project root. Claude Code reads it on every session start in that directory:

vi CLAUDE.md

Add your project instructions:

# My Project

## Build and Test
- Run tests: `npm test`
- Build: `npm run build`
- Lint: `npm run lint`

## Code Style
- Use TypeScript strict mode
- Prefer functional components in React
- All API responses must include error handling
- Never commit .env files

## Architecture
- Backend: Express.js + PostgreSQL
- Frontend: React + TailwindCSS
- Auth: JWT tokens stored in httpOnly cookies

See @README.md for project overview.
See @package.json for available scripts.

The @filename syntax imports content from other files, and Claude reads them automatically.

Path-Specific Rules

Rules that only apply to certain files. Create .claude/rules/api.md:

vi .claude/rules/api.md

Add frontmatter with path patterns:

---
paths:
  - "src/api/**/*.ts"
---

# API Rules
- All endpoints must validate input with Zod
- Return proper HTTP status codes
- Include rate limiting on public endpoints

Generate a CLAUDE.md automatically for any project with /init.

Permission Modes

Claude Code decides which tool calls need your approval based on the permission mode. With nothing configured, interactive sessions start in auto, where a classifier approves routine calls and blocks risky ones. Toggle between modes with Shift+Tab. For production lockdown configurations, see the DevOps setup guide which covers permission rules, safety hooks, and CLAUDE.md hardening for infrastructure work.

Mode	Behavior	Use Case
manual (config value default)	Prompts on first use of each tool. A grey pause badge in the footer shows when it is active	Normal interactive work
plan	Read-only: no edits, no commands	Safe exploration and planning
acceptEdits	Auto-approves file edits, prompts for commands	Trusted editing sessions
dontAsk	Auto-denies unless pre-approved	Strict control
auto	AI classifier decides allow/deny per tool call	Trusted projects with clear CLAUDE.md rules
bypassPermissions	Skips all prompts (dangerous)	Containers/sandboxes only

A cloned repo can no longer switch these risky modes on for you: "defaultMode": "auto" and "bypassPermissions" are ignored in .claude/settings.json and .claude/settings.local.json. Set them in ~/.claude/settings.json or pass --permission-mode at launch.

Pre-Approve Specific Tools

In .claude/settings.json or via the CLI:

# Allow git commands and file reads without prompting
claude --allowedTools "Read,Bash(git *)"

# Block dangerous tools explicitly
claude --disallowedTools "Bash(rm:*),Bash(sudo:*),Bash(chmod:*)"

# Combine both: allow some, block others
claude --allowedTools "Bash(git:*),Read,Edit" --disallowedTools "Bash(rm -rf:*)"

Or in settings:

vi .claude/settings.json

Add permission rules:

{
  "permissions": {
    "allow": [
      "Read",
      "Bash(git *)",
      "Bash(npm run *)",
      "Bash(docker compose *)"
    ],
    "deny": [
      "Bash(rm -rf *)"
    ]
  }
}

Deny rules are checked before allow rules, so a command that matches both lists is blocked.

Models and Effort Levels

Alias	Resolves To	Best For
opus	Claude Opus 5.5	The default model on every paid plan (Pro and Team Standard switched from Sonnet). 1M context, $4/$20 per million tokens on the API
sonnet	Claude Sonnet 5.5	Daily coding at lower cost: edits, refactors, tests. Always runs with a 1M-token context
haiku	Claude Haiku 5.5	Simple tasks, quick questions, cheap background work
fable	Claude Fable 5.2	The most capable model, for the hardest, longest-running tasks. $8/$40 per million tokens, $16/$80 in fast mode. Never a default; pick it with /model fable. Fable 5.1 and Fable 5 stay reachable as claude-fable-5-1 and claude-fable-5. In Claude apps gateway sessions the alias still means Fable 5 unless ANTHROPIC_DEFAULT_FABLE_MODEL is set. On some plans Fable bills usage credits
opusplan	Opus, then Sonnet	Opus reasoning during plan mode, then auto-switches to Sonnet for execution
opus[1m] / sonnet[1m]	1M-token context	Only matters on Opus 4.6 and Sonnet 4.6. Every Fable model, Sonnet 5.5, Sonnet 5, and Opus 4.7 and later run 1M on every plan with no suffix
default / best	Tier default / same as opus	default reverts to your account’s recommended model; best now uses the same model as opus, so it no longer reaches Fable

Switch models mid-session with /model opus or Option+P (Mac) / Alt+P (Linux/Windows). The /model picker saves your choice as the default for new sessions (press s to switch for the current session only). Aliases like opus always point to the latest release, so pin a full name (claude-opus-5-5) when you need a fixed version. Opus 5.5 needs Claude Code 2.1.280 or later, Sonnet 5.5 needs 2.1.284, the haiku alias reaches Haiku 5.5 from 2.1.293, Fable 5.2 needs 2.1.321, Fable 5.1 needs 2.1.257, and Opus 5 needs 2.1.219, so run claude update if /model does not list one. ANTHROPIC_DEFAULT_MODEL sets the model new sessions start on, but only when nothing else picks one: no --model, no ANTHROPIC_MODEL, and no model in any settings file, including one saved earlier by /model.

Effort levels control how much adaptive reasoning Claude applies. The available levels depend on the model: every Fable model, Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 5, Sonnet 5, Opus 4.8, and Opus 4.7 support all five, while Sonnet 4.6 and Opus 4.6 support all of them except xhigh (an xhigh request on those falls back to high). The default is high everywhere except Fable 5.2 and the 5.5 generation (Opus 5.5, Sonnet 5.5, and Haiku 5.5), which start at medium, and Opus 4.7, which starts at xhigh. If Fable 5.2, Opus 5.5 or Sonnet 5.5 feels shallower than the model you used before, that default is usually why.

Level	Use Case
/effort low	Short, scoped, latency-sensitive tasks that are not intelligence-sensitive
/effort medium	Routine refactors, single-file changes, test generation. Default on Fable 5.2, Opus 5.5, Sonnet 5.5, and Haiku 5.5.
/effort high	Multi-file edits, debugging, non-trivial work. Default on Fable 5.1, Fable 5, Opus 5, Sonnet 5, Opus 4.8, Opus 4.6, and Sonnet 4.6.
/effort xhigh	Design decisions, tricky cross-module refactors, gnarly bugs. Default on Opus 4.7. Available on every effort-capable model except Opus 4.6 and Sonnet 4.6 (Haiku 4.5 has no effort levels; Haiku 5.5 supports all five)
/effort max	The hardest problem of the week. Deepest reasoning, no token cap, session-only, and prone to overthinking
/effort ultracode on|off	A Claude Code switch rather than a model level: Claude orchestrates a dynamic workflow of background agents for substantive tasks. Since 2.1.284 it no longer forces xhigh and stays on at whatever effort you pick (Tab toggles it in the /effort slider)

A level typed after /effort (or confirmed with Enter in the slider) is saved as that model’s default; press s in the slider to apply it to this session only. Set effort for a single session at launch with claude --effort high, or globally with the CLAUDE_CODE_EFFORT_LEVEL environment variable, which overrides every other source. Run /effort auto to reset to the model default.

For one-off deep reasoning without changing your session effort, put the word ultrathink anywhere in your prompt. Claude Code recognizes it and adds an in-context instruction for that turn only. Note that the older “think”, “think hard”, and “think harder” phrases are no longer special keywords; they are passed through as ordinary prompt text, so ultrathink is the one to remember.

Turn On Fast Mode for Fable 5.2

Fast mode serves the same model up to 2.5 times faster at a higher token price. Claude Code 2.1.321 added it for Fable 5.2, at $16/$80 per million tokens, twice Fable’s standard rate according to Anthropic’s Fable 5.2 announcement. There is no CLI flag for it: claude --help on 2.1.322 lists nothing with “fast” in it. Start on Fable, then open the dialog:

claude --model fable

Inside the session, type /fast, press Space to toggle and Enter to confirm, or press Option+O / Alt+O to toggle it directly. A ↯ marker beside the prompt means it is on. The order matters on Fable. "fastMode": true in settings does nothing for Fable 5.2 until you run /fast while Fable 5.2 is the active model, and switching to Fable from a fast Opus session turns fast mode off unless you already turned it on while on Fable 5.2. Our lab session ran on a placeholder API key, so the availability check failed and the dialog showed this instead of the toggle:

On a real account, “Fast mode unavailable due to network connectivity issues” means Claude Code could not reach api.anthropic.com for that check. It skips ANTHROPIC_BASE_URL, so a gateway-only network trips it. Allowlist api.anthropic.com or set CLAUDE_CODE_SKIP_FAST_MODE_NETWORK_ERRORS=1. On Pro, Max and Team plans fast mode bills usage credits only, never your plan limits, so turn usage credits on first or /fast answers “Fast mode requires usage credits”. Team and Enterprise also need an Owner to enable fast mode. The first toggle in a conversation pays the uncached fast price on the whole context, so turn it on at the start of a session. The fast mode docs list the other refusal messages.

MCP Servers – Extend Claude Code

MCP (Model Context Protocol) lets Claude Code connect to external tools and data sources. Project servers live in .mcp.json at the repo root. Personal servers you want in every project are stored in ~/.claude.json, which claude mcp add --scope user writes for you. The MCP servers setup guide walks through GitHub, Postgres, and browser servers end to end.

vi .mcp.json

Example configuration with GitHub and PostgreSQL servers:

{
  "mcpServers": {
    "github": {
      "type": "stdio",
      "command": "npx",
      "args": ["@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "$GITHUB_TOKEN"
      }
    },
    "postgres": {
      "type": "stdio",
      "command": "npx",
      "args": ["@modelcontextprotocol/server-postgres"],
      "env": {
        "DATABASE_URL": "postgresql://user:pass@localhost:5432/mydb"
      }
    }
  }
}

You can also add servers from the command line instead of hand-editing JSON. Claude Code writes the config for you:

# Add an HTTP server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# Add a server to every project (user scope, stored in ~/.claude.json)
claude mcp add --transport http --scope user sentry https://mcp.sentry.dev/mcp

# Add a stdio server with environment variables
claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=$GITHUB_TOKEN -- npx @modelcontextprotocol/server-github

# List, inspect, or remove a server
claude mcp list
claude mcp get github
claude mcp remove github

# Log in or out of a server that uses OAuth
claude mcp login github
claude mcp logout github

Manage MCP connections with /mcp in the interactive session or pass a config file on the CLI:

# Load MCP servers from an external config
claude --mcp-config ./team-mcp.json

# Load multiple configs
claude --mcp-config ./github-mcp.json ./db-mcp.json

# Ignore all other MCP sources, only use this config
claude --mcp-config ./ci-mcp.json --strict-mcp-config

The --mcp-config flag is useful in CI/CD where you need consistent MCP server configuration across runs. --strict-mcp-config ensures only the specified servers are loaded, ignoring any from .mcp.json or user settings.

Custom Skills – Your Own Slash Commands

Skills are reusable prompts that show up as slash commands. Create them at .claude/skills/ (project) or ~/.claude/skills/ (personal).

Example, create a deploy skill:

mkdir -p .claude/skills/deploy
vi .claude/skills/deploy/SKILL.md

Add the skill definition:

---
name: deploy
description: Deploy the application to production
argument-hint: "[environment]"
user-invocable: true
---

Deploy the application to the $0 environment (default: staging).

Steps:
1. Run the test suite
2. Build the production bundle
3. Deploy using the deploy script
4. Verify the deployment health check

Now use it: /deploy production

Custom Subagents

Subagents are specialized AI agents that Claude can delegate tasks to. Each one is a single Markdown file in .claude/agents/ (project) or ~/.claude/agents/ (personal). The old /agents wizard is gone: typing it now only reminds you to ask Claude to write the file or edit it yourself. The subagents guide covers tool scoping and model choice in more depth.

mkdir -p .claude/agents
vi .claude/agents/code-reviewer.md

Define the agent:

---
name: code-reviewer
description: Expert code reviewer. Use proactively after code changes.
tools: Read, Grep, Glob, Bash
model: sonnet
---

You are a senior code reviewer. When reviewing code:
- Check for security vulnerabilities (SQL injection, XSS, etc.)
- Verify error handling is complete
- Flag any hardcoded secrets or credentials
- Suggest performance improvements
- Check for proper input validation

Claude automatically delegates code review tasks to this agent, or you can invoke it directly with @code-reviewer check my latest changes. Add omitClaudeMd: true to the frontmatter when a narrow agent should run without your user, project, and local CLAUDE.md files; it saves context on every spawn.

From the CLI, select an agent for the entire session:

# Use a named agent from .claude/agents/
claude --agent code-reviewer

# In print mode (CI/CD)
claude -p "review the latest commit" --agent code-reviewer

# Define agents inline (no files needed)
claude --agents '{"qa": {"description": "QA tester", "prompt": "You are a QA engineer. Write test cases for every change."}}'

Agents passed with --agents exist only for that session. The files in .claude/agents/ are the ones to commit so the whole team gets them.

Hooks – Automation Triggers

Hooks run shell commands automatically in response to Claude Code events. Configure in .claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "npx prettier --write $(jq -r '.tool_input.file_path')"
          }
        ]
      }
    ],
    "Notification": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "osascript -e 'display notification \"Claude needs input\" with title \"Claude Code\"'"
          }
        ]
      }
    ]
  }
}

The events you will reach for most are SessionStart, UserPromptSubmit, PreToolUse, PermissionRequest, PostToolUse, Stop, SubagentStop, PreCompact, Notification, and SessionEnd. Newer ones include PostToolUseFailure, FileChanged, CwdChanged, DirectoryAdded (fires after /add-dir), WorktreeCreate, and PreModelSwitch. The Claude Code hooks guide has working blockers and formatters for each, and the SSH server management guide shows hooks enforcing safety on production servers.

Context Window Management

Claude Code has a finite context window. Managing it well is the difference between a productive session and hitting walls.

Key Commands

Command	When to Use
/compact	Summarize conversation to free space mid-task
/compact Focus on auth module	Summarize but keep specific context intact
/clear	Complete reset between unrelated tasks
/context	See what is consuming your context

Tips for Long Sessions

• Use /clear between unrelated tasks, so old context never leaks into new work
• Run /compact proactively when you notice slowdowns
• Use subagents for verbose operations (test runs, log analysis); their output stays in the subagent context
• Move detailed instructions to skills instead of typing them every time
• Keep CLAUDE.md under 200 lines. Move details to separate files and import with @filename

Git Worktrees – Parallel Work

Work on multiple tasks simultaneously without branch switching conflicts:

# Start Claude in a new worktree
claude -w feature-auth

# Batch changes across many files in parallel
/batch migrate all components from class-based to functional

The -w flag creates an isolated git worktree. Changes happen in the worktree without affecting your main working directory. When the agent finishes, merge the worktree branch back.

Background Agents and Dynamic Workflows

Background agents are full Claude Code sessions that run detached from your terminal, so you can fire off a long task and keep working in the foreground. They keep running after you close the terminal and are managed from a small set of subcommands. This is the largest addition to Claude Code since the cheat sheet first shipped.

Start one from the shell with --bg. It prints a session ID and returns immediately:

# Launch a background agent and get its ID back
claude --bg "investigate the flaky test in tests/auth"

# Run a shell command as a background job instead of a session
claude --bg --exec 'pytest -x'

# Run a specific subagent in the background
claude --bg --agent code-reviewer "review the open PR"

# Background agents can commit, push, and open a draft PR when they finish
claude --bg "fix the failing CI checks, then commit and open a PR"

From inside an interactive session, detach the whole thing with /background (alias /bg) and list what is running with /tasks. /fork works the other way around: it copies the current conversation into a new background session and leaves you working in the original, so a side investigation can run with full context without stealing your terminal (reach for /subtask when you want that fork to run inside the session instead). To watch and dispatch background sessions in a live view, open the agent manager:

claude agents

Manage individual background sessions from the shell by ID. The agent view shows the IDs, or you can list them as JSON for scripting:

# List live background sessions as JSON
claude agents --json

# Attach a background session to this terminal
claude attach 7c5dcf5d

# Tail its recent output
claude logs 7c5dcf5d

# Stop, restart (keeps the conversation), or remove from the list
claude stop 7c5dcf5d
claude respawn 7c5dcf5d
claude rm 7c5dcf5d

Background sessions run under a supervisor process. If the agent view reports that the background service did not respond, inspect or reset it:

# Show supervisor state, version, and worker count
claude daemon status

# Stop the supervisor but keep workers so the next one reconnects
claude daemon stop --any --keep-workers

Dynamic workflows take this further. A single instruction is decomposed into many background subagents that run in parallel. Trigger one for a substantive task by setting the ultracode effort level, then run /workflows to watch progress live. By default Claude aims for a medium workflow (fewer than 10 agents), or small (fewer than 5) on Pro plans. The guideline is advice rather than a hard cap; change it or remove it under Dynamic workflow size in /config, or with the workflowSizeGuideline setting. For a codebase-wide change, /batch splits the work into independent units and runs each in its own git worktree with its own pull request.

The 200-per-session subagent spawn cap was removed in 2.1.224, and CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION is now a no-op; concurrency and depth limits still apply. Web searches are still capped at 200 per session (CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION). Any MCP tool call that runs past two minutes moves to the background automatically so the session stays usable (CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS).

Want a leaner session? Set CLAUDE_CODE_DISABLE_AGENT_VIEW=1 to turn off background agents and CLAUDE_CODE_DISABLE_WORKFLOWS=1 to turn off dynamic workflows.

If you run a long-lived agent gateway next to Claude Code, the OpenClaw cheat sheet (/openclaw-cheat-sheet/) covers its Gateway service, cron automations and the commands other sheets get wrong.

Settings File Reference

File	Scope
.claude/settings.json	Project (shared via git)
.claude/settings.local.json	Project (gitignored, personal)
~/.claude/settings.json	All projects (personal)

Common settings:

{
  "model": "opus",
  "modelSettings": {
    "claude-opus-5-5": { "effortLevel": "high" }
  },
  "autoMemoryEnabled": true,
  "permissions": {
    "defaultMode": "manual",
    "allow": ["Read", "Bash(git *)"],
    "deny": ["Bash(rm -rf *)"]
  }
}

Effort is saved per model under modelSettings, which is what /effort high writes for you. A top-level "effortLevel" in ~/.claude/settings.json no longer counts for Opus 5.5, so the older form silently leaves it at medium. A few newer keys worth knowing: "attribution": false hides the commit and PR attribution lines (older CLI versions skip a settings file that holds it, so keep the object form in files shared across versions), modelPicker curates the /model list, and maxProseWidth caps prose width in wide terminals while tables and code keep the full width.

Gitignoring settings.local.json also makes it a quiet leak path: nothing flags what a “don’t ask again” approval stores in it, and an approval keeps the literal command, token included. Before you copy the file to another machine or hand it to a teammate, audit settings.local.json for leaked secrets, then pull out any allow entry you cannot explain.

Environment Variables

Variable	Purpose
ANTHROPIC_MODEL	Force the session model (e.g., opus); wins over a saved /model pick
ANTHROPIC_DEFAULT_MODEL	Model new sessions start on when no flag, ANTHROPIC_MODEL, or saved model setting picks one
ANTHROPIC_API_KEY	API key for direct API access
CLAUDE_CODE_EFFORT_LEVEL	Default effort (low/medium/high/xhigh/max)
CLAUDE_CODE_DISABLE_AUTO_MEMORY	Set to 1 to disable auto-memory
CLAUDE_CODE_DISABLE_1M_CONTEXT	Set to 1 to disable 1M token context
MAX_THINKING_TOKENS	Custom thinking budget (default varies)
HTTPS_PROXY	HTTP proxy for corporate networks
CLAUDE_CODE_NO_FLICKER	Set to 1 to use the fullscreen TUI renderer (same as tui: "fullscreen"): smooth, flicker-free output with mouse support
ENABLE_TOOL_SEARCH	Controls MCP tool search. Unset or true defers MCP tools, including behind an ANTHROPIC_BASE_URL gateway since 2.1.304; auto turns it on once tool definitions reach 10% of context (auto:N sets another threshold), and false loads every tool upfront
CLAUDE_CODE_SUBAGENT_MODEL	Default model for subagents, teammates, and workflow agents. A model named at spawn time or in the agent file wins; add CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 to force it everywhere
CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION	Cap on WebSearch tool calls per session (default 200)
CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS	How long an MCP call runs before moving to the background (default 2 minutes)
CLAUDE_CODE_DISABLE_AGENT_VIEW	Set to 1 to turn off background agents
CLAUDE_CODE_DISABLE_WORKFLOWS	Set to 1 to turn off dynamic workflows
CLAUDE_CODE_DISABLE_BACKGROUND_TASKS	Set to 1 to disable background bash tasks
CLAUDE_CODE_TASK_LIST_ID	Share a task list across sessions via a named directory in ~/.claude/tasks/
DISABLE_AUTOUPDATER	Set to 1 to disable automatic updates entirely
ANTHROPIC_DEFAULT_FABLE_MODEL	Pin which model the fable alias resolves to
CLAUDE_CODE_AUTO_COMPACT_WINDOW	Token threshold to auto-compact before the window fills, from 100k to 1M: a plain integer or, since 2.1.304, a suffixed value such as 200k or 1M. Out-of-range values are clamped with a warning
CLAUDE_CODE_RETRY_WATCHDOG	Set to 1 for CI or unattended runs: retries 429 and 529 capacity errors indefinitely (backing off up to 5 minutes) instead of failing after CLAUDE_CODE_MAX_RETRIES. A 429 that reports a spend limit or exhausted credits still fails at once
CLAUDE_CODE_RESTRICTED	Set to 1 to start in restricted mode, same as --restricted
CLAUDE_CODE_SAFE_MODE	Set to 1 to start with every customization disabled, same as --safe-mode
CLAUDE_CODE_AUTO_MODE_SERVER	Set to 0 to keep auto mode’s classifier local instead of the server-side reviewer
CLAUDE_CODE_AUTO_MODE_MODEL	Model ID for auto mode’s classifier requests
CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS	Context size where idle compaction applies (default 200000)
CLAUDE_CODE_DISABLE_WEB_FETCH	Set to 1 to turn off WebFetch while keeping WebSearch
CLAUDE_CODE_DISABLE_FILE_MENTIONS	Set to 1 so @ mentions pass paths instead of expanding file contents

Two variables without their own row in the table decide where requests go and how they authenticate. ANTHROPIC_BASE_URL points Claude Code at an LLM gateway instead of the Anthropic API, and ANTHROPIC_AUTH_TOKEN carries the gateway credential as a Bearer header. Pair them with ANTHROPIC_MODEL and you can run Claude Code on free-tier models behind a self-hosted FreeLLMAPI proxy. Anthropic does not support routing Claude Code to non-Claude models through a gateway, so treat that setup as a cost-saving experiment rather than your daily driver.

The CLAUDE_CODE_NO_FLICKER variable is worth calling out. By default, the terminal renderer redraws the entire output with each token, causing visible flashing in code blocks and tables. With CLAUDE_CODE_NO_FLICKER=1, output streams line-by-line without screen jumps, syntax highlighting stays stable, and table columns stay aligned while populating. It also enables mouse event support (clicking and scrolling). Add it to your shell profile for a noticeably smoother experience:

echo 'export CLAUDE_CODE_NO_FLICKER=1' >> ~/.zshrc
source ~/.zshrc

Or use it in an alias that bundles your preferred defaults:

alias c="CLAUDE_CODE_NO_FLICKER=1 claude --effort max"

This is the same renderer you can turn on permanently with "tui": "fullscreen" in settings. The fullscreen view also unlocks transcript-viewer shortcuts: press ? for the full shortcut panel and { / } to jump between prompts.

File Structure Reference

Claude Code reads configuration from two scopes: your home directory for personal defaults and the project root for repo-specific rules. The tree below maps the files you’ll touch most:

# User-level (personal, all projects)
~/.claude.json           # User-scoped MCP servers, app state
~/.claude/
  settings.json          # Global settings
  keybindings.json       # Custom keyboard shortcuts
  CLAUDE.md              # Personal instructions
  rules/                 # Personal rules
  agents/                # Personal subagents
  skills/                # Personal skills

# Project-level (shared with team)
.claude/
  settings.json          # Project settings (git tracked)
  settings.local.json    # Local overrides (gitignored)
  CLAUDE.md              # Project instructions
  rules/                 # Path-specific rules
  agents/                # Project subagents
  skills/                # Project skills
.mcp.json                # Project MCP servers
CLAUDE.md                # Project instructions (alternative location)
AGENTS.md                # Read when no CLAUDE.md exists

Everything in the project half is meant for git except .claude/settings.local.json, which holds your personal overrides and should stay out of the repo.

Practical Workflow Tips

Start Every Project Right

The two-minute ritual that makes every subsequent session faster: seed a CLAUDE.md at the project root so Claude knows your conventions without being told each time:

# In your project root
claude
/init

This scans your project and generates a CLAUDE.md with build commands, test instructions, and code conventions. Review and edit it, because this file shapes every future interaction.

Plan Before You Build

Press Shift+Tab to enter plan mode before asking Claude to make changes. In plan mode, Claude reads and analyzes but cannot edit anything. Once you approve the plan, switch back to default mode and say “implement the plan”.

Undo Mistakes Instantly

Press Esc + Esc to open the rewind menu. Choose “Restore code and conversation” to undo both the code changes and the conversation that led to them. This is faster than git stash for quick experiments.

Pipe Everything

Claude Code is a normal Unix citizen on stdin/stdout. Anything you can pipe into grep, you can pipe into claude -p:

# Analyze logs
tail -500 /var/log/app.log | claude -p "find the root cause of errors"

# Review a PR
gh pr diff 42 | claude -p "security review this PR"

# Generate commit message
git diff --staged | claude -p "write a conventional commit message" --model haiku

# Translate
cat README.md | claude -p "translate to Japanese" > README.ja.md

# Docker container analysis
docker ps --format json | claude -p "which containers are using the most resources?"

# Kubernetes pod debugging
kubectl logs deployment/api --tail=200 | claude -p "why are requests failing?"

# Database structure review
pg_dump --schema-only mydb | claude -p "suggest index improvements"

# Summarize recent commits
git log --oneline -20 | claude -p "write release notes from these commits"

# Find config issues
cat /etc/nginx/nginx.conf | claude -p "check for misconfigurations"

For deeper walkthroughs of Docker and Kubernetes workflows with Claude Code, see the Docker guide and Kubernetes guide.

Automation Scripts

Chain multiple claude -p calls in a bash script for repeatable workflows:

#!/bin/bash
# Automated PR review pipeline
git diff HEAD~1 > /tmp/diff.txt
cat /tmp/diff.txt | claude -p "review for security issues" --model sonnet > security_review.md
cat /tmp/diff.txt | claude -p "check for performance regressions" --model sonnet > perf_review.md
cat /tmp/diff.txt | claude -p "write a one-paragraph summary" --model haiku > summary.md

Batch process multiple files:

#!/bin/bash
# Generate docstrings for all Python files missing them
for f in $(grep -rL '"""' src/*.py); do
  claude -p "add docstrings to functions missing them in $f" \
    --allowedTools "Read,Edit" --max-turns 3
done

Generate a changelog from git tags:

git log v1.0.0..v2.0.0 --oneline | \
  claude -p "create a changelog grouped by feature, fix, and chore" \
  --output-format text > CHANGELOG.md

Text is already the default output format for -p, so the flag is there for whoever reads the script next: it makes plain that CHANGELOG.md gets prose, not JSON.

Side Questions with /btw

Need to ask something without derailing the current task? Use /btw:

/btw what was the name of that config file we edited earlier?

The response is ephemeral (not saved to context), low-cost (uses cache), and does not use any tools.

Cost Control

• Use /cost to check spend at any time; it now shows the prompt-cache hit ratio and a likely cause for each cache miss
• Use Sonnet for most work, Opus only for complex problems
• Use /effort low for simple tasks
• Run /compact to reduce context size (fewer tokens per request)
• Be specific in prompts, because vague requests cause more back-and-forth
• Use --max-budget-usd 5.00 in CI/CD to prevent runaway costs

Prompting Tips – Get Better Results

The quality of Claude Code’s output depends heavily on how you prompt it. These patterns consistently produce better results.

Be Specific, Not Vague

Bad Prompt	Good Prompt
“fix the bug”	“the login endpoint returns 500 when email contains a plus sign; fix the input validation in src/auth/login.ts”
“make it faster”	“the /api/users endpoint takes 3s on 10k rows; add database indexing and pagination”
“write tests”	“write unit tests for the calculateDiscount function in src/pricing.ts covering edge cases: zero price, negative quantity, expired coupon”
“clean up the code”	“extract the duplicate validation logic in src/api/orders.ts and src/api/returns.ts into a shared validator”

Use Plan Mode First

For complex tasks, switch to plan mode (Shift+Tab) and ask Claude to explore before coding:

# In plan mode:
"I need to add OAuth2 login with Google. Look at the existing auth flow
in src/auth/ and propose where to add the Google provider without
breaking the current email/password login."

Review the plan, then switch back to default mode and say “implement the plan”.

Give Verification Targets

Tell Claude how to verify its own work:

"Add rate limiting to the /api/auth endpoints. After implementing,
run npm test to make sure nothing breaks, and verify the rate limiter
works by checking the middleware is registered in the Express app."

Naming the check gives Claude a pass or fail signal to work toward, so it does not stop the moment the code compiles.

Use @file Mentions for Context

Point Claude at specific files instead of making it search:

"look at @src/db/schema.ts and @src/api/users.ts - the User model
has a new 'role' field in the schema but the API doesn't expose it.
Add the role field to the GET /users/:id response."

Each @ mention pulls that file straight into context, which saves the search turns and the tokens they would cost.

Course-Correct Early

If you see Claude going in the wrong direction, press Ctrl+C to stop and redirect. Do not wait for it to finish a large wrong change. Interrupt and clarify. If it already made bad changes, press Esc + Esc to rewind.

Break Large Tasks Into Steps

Instead of “build a user management system”, split it:

1. “Create the User database schema with id, email, name, role, created_at”
2. “Add CRUD API endpoints for users with input validation”
3. “Write tests for the user endpoints”
4. “Add role-based access control middleware”

Each step gets verified before moving to the next. This avoids wasting tokens on a bad foundation.

Token Saving Strategies

Every message sent to Claude includes the full conversation context. Larger context means higher cost and slower responses. These strategies keep your sessions efficient.

Context Hygiene

Strategy	How	Savings
Clear between tasks	/clear when switching to unrelated work	Biggest impact: resets to zero
Compact proactively	/compact Focus on the auth module changes	Keeps relevant context, drops noise
Check context usage	/context to see what is consuming space	Identifies bloat sources
Use @file mentions	@src/auth.ts instead of “find the auth file”	Avoids exploratory file reads
Use subagents	Delegate verbose tasks (test runs, log analysis)	Output stays in subagent context

Model Selection

Task	Model	Effort
Simple edits, renames, formatting	Sonnet	/effort low
Standard coding, tests, refactoring	Sonnet	/effort medium
Complex debugging, architecture	Opus	/effort high
Quick questions	Haiku (via /btw)	N/A

CLAUDE.md Optimization

• Keep the main CLAUDE.md under 200 lines, since it loads on every message
• Move detailed instructions into skills, which only load when invoked
• Use @file imports for large reference docs instead of pasting content
• Put path-specific rules in .claude/rules/, which only load when Claude touches matching files

MCP and Tool Optimization

• Disable unused MCP servers with /mcp, since each server adds tool definitions to context
• Use CLI tools (gh, aws, gcloud) instead of MCP when a single command works
• Leave MCP tool search on (the default): tool definitions load on demand instead of all at once. Run /skill-doctor and claude plugin details <name> to find skills and plugins that cost context without earning it

CI/CD Budget Controls

Any Claude Code invocation in CI should have a dollar cap and a fallback model, which matters more now that the top tier, Claude Fable 5.2, costs $8/$40 per million tokens and $16/$80 with fast mode on. Skip these and one bad prompt can drain your monthly budget in a single failing pipeline:

# Hard spending limit
claude -p "fix failing tests" --max-budget-usd 2.00

# Limit agentic turns
claude -p "review this file" --max-turns 3

# Use cheapest model for simple CI tasks
claude -p "check for linting errors" --model haiku

--max-turns exits with an error at the limit, so a CI job stuck in a loop fails fast. --max-budget-usd checks a client-side cost estimate and spend can pass it slightly, so leave headroom.

Diagnostics – /doctor and /insights

Three tools for seeing exactly what Claude Code is doing during a session:

–debug – Verbose Logging

When you need to see exactly what Claude Code is doing (API calls, tool invocations, timing), use --debug:

# Full debug output to stderr
claude --debug

# Filter to specific categories
claude --debug "api,hooks"

# Write debug logs to a file (keeps terminal clean)
claude --debug-file /tmp/claude-debug.log

This is especially useful when MCP servers are not connecting, hooks are not firing, or API calls are failing silently.

/doctor and /checkup – Full Setup Checkup

Run /doctor, or its alias /checkup, when something is not working right or when you want to tidy up your setup. It is no longer a read-only report. It runs a full checkup, and where it finds problems you press f to have Claude fix them. It confirms with you before changing anything. The checkup covers:

• Claude Code version and auto-updater status, and updating to the latest release
• Authentication state, token validity, and API connectivity
• Settings file syntax and conflicts
• MCP server health, plus removing unused skills, MCP servers, and plugins to save context
• De-duplicating your personal CLAUDE.md against the project one, and splitting an overgrown root CLAUDE.md into nested files and skills
• Turning off slow hooks, enabling auto mode, and pre-approving frequently denied read-only commands

/insights – Session Analysis

Run /insights after a long session to get an analysis report covering:

• Token usage breakdown by category
• Most expensive operations in the session
• Patterns in how you use Claude Code
• Suggestions for improving efficiency
• Time spent waiting vs. actively working

Use this to identify habits that waste tokens and adjust your workflow.

IDE Integrations

Claude Code integrates with VS Code and JetBrains IDEs for inline diffs, context sharing, and side-by-side conversations. Wondering how it stacks up against other AI coding tools? See the OpenCode vs Claude Code vs Cursor comparison.

VS Code

• Install the “Claude Code” extension from the VS Code marketplace
• Open Command Palette (Cmd+Shift+P on Mac / Ctrl+Shift+P on Linux/Windows) and search “Claude Code”
• Works with Cursor editor too

JetBrains (IntelliJ, PyCharm, WebStorm, etc.)

• Install “Claude Code” from the JetBrains Marketplace
• Works across all JetBrains IDEs
• Supports interactive diffs and context sharing

Start Claude with IDE auto-detection:

claude --ide

With --ide, Claude Code connects on startup when exactly one supported IDE is running; with several open, pick one from /ide instead.

Auto Mode

Auto mode uses an AI classifier to decide whether each tool call needs your approval. Instead of prompting on every command, it allows safe operations (reads, git status, test runs) and blocks risky ones (force push, rm -rf, production deploys) automatically. Enable it with --permission-mode auto or press Shift+Tab to cycle to it.

Inspect the classifier rules with the claude auto-mode subcommand:

# View current auto-mode config
claude auto-mode config

# View default allow/deny rules
claude auto-mode defaults

# Get AI feedback on your custom rules
claude auto-mode critique

Customize rules in .claude/settings.json or ~/.claude/settings.json under the autoMode key with allow, soft_deny, and environment arrays. The classifier reads these plus the defaults to make per-tool-call decisions. The same rules are editable in the Auto mode tab of /permissions, and /insights now estimates how many of your recent permission prompts auto mode would have handled.

/loop – Recurring Tasks

The /loop slash command runs a prompt or another slash command on a repeating interval within your session. Useful for polling deploy status, watching test results, or checking a service health endpoint.

# Check deploy status every 5 minutes
/loop 5m check if the deploy on staging is complete

# No interval: Claude picks the pace between iterations
/loop /test

# Custom interval with a slash command
/loop 2m /security-review

The loop runs in the background until you stop it or end the session. Each iteration gets its own context, so long-running loops do not bloat the conversation.

Subcommands Reference

Claude Code provides subcommands for management tasks outside of interactive sessions:

Subcommand	What It Does
claude agents	Open the agent view to monitor and dispatch parallel background sessions (--json for scripting)
claude attach <id>	Attach a background session to this terminal
claude logs <id>	Print recent output from a background session
claude stop <id>	Stop a background session (also claude kill)
claude respawn <id>	Restart a background session with its conversation intact
claude rm <id>	Remove a background session from the list (transcript stays on disk)
claude daemon status	Show the background-session supervisor state and worker count
claude auth	Manage authentication (login, logout, status)
claude auto-mode config	Print effective auto-mode classifier rules
claude auto-mode defaults	Show default allow/deny rules
claude auto-mode critique	Get AI feedback on your custom auto-mode rules
claude auto-mode reset	Restore the default auto-mode configuration (add --yes to skip the confirmation)
claude doctor	Health check for auto-updater, auth, MCP, settings
claude import [codex|gemini|cursor]	Import instruction files, MCP servers, commands, subagents, and skills from another agent (--dry-run first)
claude install [target]	Install a specific version (stable, latest, or a version number)
claude mcp	Configure and manage MCP servers from the CLI
claude plugin	Manage plugins (alias claude plugins): install, list --json, details (token cost), validate --json, init
claude plugin eval [target]	Run a plugin’s eval suite against Claude Code and get scored results as JSON plus an HTML report
claude purge [path]	Delete all local Claude Code state for a project (transcripts, tasks, history). --dry-run previews, --all covers every project. Was claude project purge
claude remote-control	Run a Remote Control server to drive Claude Code from claude.ai or the Claude app
claude ultrareview [target]	Run a cloud-hosted multi-agent code review and print the findings (--json, --timeout)
claude setup-token	Set up a long-lived auth token (requires subscription)
claude update	Check for and install updates

Claude Code for DevOps – Full Guide Series

This cheat sheet covers the commands and configuration. The DevOps series goes deeper with real-world demos, tested on actual infrastructure:

• Set Up Claude Code for DevOps Engineers – CLAUDE.md rules, permission lockdowns, hooks, and safety configuration for production use
• Manage Servers with Claude Code via SSH – multi-server health checks, log analysis, config audits, and incident response over SSH
• Build and Debug Docker Containers with Claude Code – Dockerfile generation, multi-stage builds, Compose workflows, and container debugging
• Deploy Infrastructure with Claude Code and Terraform – HCL generation, plan review, state management, and module refactoring
• Generate and Debug Ansible Playbooks with Claude Code – playbook generation, role scaffolding, multi-OS support, and Vault integration
• Deploy and Debug Kubernetes Apps with Claude Code – manifest generation, pod debugging, Helm chart management, and cluster troubleshooting
• Set Up Claude Code in GitHub Actions for Infrastructure Review – automated PR reviews, Terraform plan analysis, and CI/CD integration

If you would rather learn the workflow from a book than from release notes, Build with Claude Code (paid link) by Taeho Jo (Packt, August 2026, 408 pages) walks from a first session to deployed apps and is the most hands-on of the seven titles in our AI coding books round-up. Books lag the CLI by a few releases, so keep this page open next to it for the current flags and shortcuts.

Claude Code Alternatives for Other Models

Claude Code only talks to Anthropic models, directly or through Bedrock, Vertex, and Foundry. When a project needs a local model or a different provider, OpenCode covers more than 75 providers, Aider keeps every change as a git commit, and Claw Code is an open-source rewrite in Rust. For an open-source agent that runs on any OpenAI-compatible endpoint, including local Ollama, see the Hermes Agent cheat sheet. If your team already pays for GitHub Copilot, its terminal agent runs Claude or GPT models (pick one with --model at launch or /model mid-session) and still loads your CLAUDE.md, and the GitHub Copilot CLI cheat sheet covers its commands, permission flags, and MCP setup. Moving the other way is easier now: claude import codex (or gemini, cursor) brings instruction files, MCP servers, and skills across, and --dry-run shows the plan before anything is written. The official Claude Code docs remain the reference for anything that changed after October 2026.

Keep reading