Chapter 14

Cheat sheet

Print it. Tape it next to your monitor. Stop re-googling.

Jump to section tap to open

Daily slash commands

/help List everything available
/init Generate starter CLAUDE.md
/clear Reset context, keep session
/compact Summarize history, free space
/cost Running spend this session
/model Switch models mid-session
/agents Manage subagent definitions
/mcp List MCP servers
/hooks See/edit hooks
/plugins Browse plugin marketplace
/install-ide Wire CC into your editor
/resume Pick a past session
/exit Leave session (auto-saved)

CLI flags

claude Interactive in current dir
claude --print "..." Headless, prints answer, exits
claude --model opus Pin a model for this session
claude --resume Pick from past sessions
claude --continue Resume most recent session
claude --plan Describe-only mode, no edits
claude --no-update Skip auto-update check
claude --version Sanity check install
claude --output-format json Parse-friendly output
claude --allowed-tools "..." Launch-time allow-list

Settings.json keys

model "sonnet""opus""haiku"
theme "dark""light""auto"
telemetry truefalse
autoUpdater "weekly""daily""disabled"
permissions.allow glob patterns to auto-approve
permissions.deny glob patterns to block
permissions.disableBypassPermissionsMode "disable"

File paths

~/.claude/settings.json Global preferences
~/.claude/commands/ Custom slash commands
~/.claude/agents/ Custom subagent definitions
~/.claude/skills/ Personal skills
~/.claude/plugins/ Installed plugin bundles
<repo>/CLAUDE.md Project memory, every turn
<repo>/.mcp.json MCP servers (commit it)
<repo>/.claude/settings.json Repo overrides
<repo>/.claude/agents/ Repo-shared subagents
<repo>/.claude/skills/ Repo-shared skills

Environment vars

ANTHROPIC_API_KEY Required for headless / CI
CLAUDE_CONFIG_DIR Relocate ~/.claude (sandboxes)
CLAUDE_MODEL Per-shell default
MCP_SERVER_<NAME>_<KEY> Forwarded into MCP processes

Keyboard shortcuts

Ctrl-C Interrupt generation, keep session
Ctrl-D / /exit Leave session
Recall last prompt
Esc Esc Undo last input
Shift-Enter Newline without send
Tab Autocomplete file paths
/ Slash-command picker
@ File-reference picker
# <text> Append to CLAUDE.md inline

Hook events (90% rule)

PreToolUse Before tool call. Validate, block.
PostToolUse After success. Format, lint, test.
Stop Turn ends. Commit, ship, notify.
SessionStart Inject context on session boot
UserPromptSubmit Before model sees a prompt

Cron syntax

0 7 * * 1-5 7 AM every weekday
*/15 * * * * Every 15 minutes
0 16 * * 5 4 PM every Friday
0 9 * * 1 9 AM every Monday
0 17 * * 1-5 5 PM ET weekdays (set TZ)
crontab.guru Translate human ↔ cron

Permission rule syntax

Bash(npm test*) Glob inside parens auto-approves
Edit(src/**/*) Recursive path glob
Edit(src) Matches one literal file "src"
WebFetch Bare name = every invocation
Edit(.env*) Common deny: secrets
Bash(rm -rf*) Common deny: destructive
bash(...) Wrong — names case-sensitive
deny → ask → allow Eval order; deny always wins
~/.ssh/** Deny follows symlinks both ways

Subagent frontmatter

name Subagent identifier
description What orchestrator reads to delegate
tools Allow-list, e.g. Read, Grep, Glob
(omit Edit/Write) Makes it physically read-only
<body> The system prompt for the job
~/.claude/agents/ User-level, every project
.claude/agents/ Repo-level, commit it
general-purpose Built-in, all tools
Explore Built-in, read-only repo search
Plan Built-in, returns plan only

SKILL.md frontmatter

name Skill identifier
description Trigger — literal phrases, exclusions
allowed-tools Optional tool allow-list
SKILL.md Only required file
scripts/ Deterministic code, optional
templates/ Output shapes to match
reference/ Long detail pulled on demand
Do NOT use for Negative space stops misfires
<40 line body Past that, model skips steps

.mcp.json shape

mcpServers Top-level object of servers
command "npx""python" launcher
args Array, e.g. ["-y", "pkg", path]
env Per-server vars (never raw keys)
stdio Local subprocess, zero network
streamable-http Hosted SaaS transport
SSE Legacy — being phased out
/mcp List + confirm connected
<repo>/.mcp.json Commit it, team shares servers

Hook JSON shape

matcher Regex vs tool name, "Edit|Write"
type: command Hook runs a shell command
command The shell line to execute
$CLAUDE_FILE_PATH File the tool touched
$CLAUDE_PROJECT_DIR Repo root
$CLAUDE_TOOL_INPUT Raw tool input for grep
exit 0 Tool call proceeds
exit non-zero Tool call blocked
stderr Surfaces to model as error
stdout Surfaces to model as context

Headless / CI one-liners

-p / --print Headless, stdout, exit code
--output-format json Single result object
--output-format stream-json NDJSON live events
total_cost_usd JSON field — spend per run
--allowed-tools "..." Launch-time allow-list
--disallowed-tools "..." Launch-time block-list
--resume <id> Reattach prior session
timeout 300 claude -p Hard cap on a loop
ANTHROPIC_API_KEY Server/CI auth (not OAuth)

Plan → Auto → /goal ladder

Plan mode Removes per-step surprise
Auto mode Removes per-tool prompts
/goal <cond> Removes per-turn prompts
/goal clear Aliases: stop off reset cancel
or stop after N Always cap blast radius
/loop 5m <prompt> Interval-driven polling
Stop hook Script-driven, settings-scoped
Haiku 4.5 Default /goal evaluator
one goal/session New goal replaces old

Model routing + cost

Haiku Triage/classify — ~1/15 Sonnet
Sonnet Default working model
Opus Hard reasoning, agent loops
output tokens ~5x input price
cache write 1.25x input rate, one-time
cache read ~0.1x input rate (10x cheaper)
5-min eviction Ephemeral cache cold window
append-only New context at bottom, not top
Batch API 50% off, returns within 24h
Stay close

The next edition lands when this list says it does.

No course. No paywall. Operator playbooks weekly. 10K+ subscribers.