What this course’s lessons - steering sessions, setting permission modes, delegating to subagents - use most: commands, shortcuts, flags, modes, files, in one place to look up rather than re-read.
| Command |
What it does |
/clear |
Start a new conversation in the same process; the previous session remains resumable |
/compact |
Summarize the thread to reclaim tokens without losing the plot |
/context |
Show what’s currently filling the context window, including which memory files loaded |
/usage |
Show token spend for this session |
/rewind |
Restore or summarize an eligible checkpoint (a saved snapshot of the session’s tracked edits); it is not a Git replacement |
/plan |
Plan a single prompt without switching the session’s mode |
/init |
Generate or improve a starter CLAUDE.md by scanning the project |
/memory |
List, open, or edit memory files (CLAUDE.md, rules) and manage auto memory (the agent’s own notes on your machine) |
/model |
Inspect models; picking one in the list (Enter) or naming it (/model <name>) saves your default, s uses the pick for this session only |
/effort |
Set supported reasoning effort levels |
/permissions |
View and manage allow, ask, and deny rules (saved to settings) |
/tasks |
List background sessions and subagents; attach to or stop each one |
/btw |
Ask a side-question in an overlay that stays out of conversation history; the newest twenty exchanges replay as context for each ask, x clears them (interactive-mode reference) |
/resume |
Pick a previous session to continue |
/branch |
Create a new conversation branch from the current session |
/fast |
Toggle fast mode, or set it with on / off; currently Opus-only (as of 2026-10; see the fast-mode reference), subscription use draws from usage credits, and it is disabled by default for Team and Enterprise organizations until an Owner enables it |
/recap |
Summarize the current session without compacting it |
/config |
Open the interactive settings panel (editor mode, recaps, theme) |
/terminal-setup |
Install the Shift+Enter binding for common terminals |
/keybindings |
Configure interactive keybindings |
| Flag |
What it does |
--continue |
Resume the most recent session in this directory |
--resume |
Pick a past session to resume from a list |
-p |
Run a single prompt headlessly and exit (no REPL) |
--output-format |
Control headless output: text, json, stream-json |
--max-turns |
Cap how many agentic turns a headless run may take |
--bare |
Skip project auto-discovery such as CLAUDE.md, hooks, skills, plugins, and MCP (the Model Context Protocol, an open standard for connecting an agent to an external tool or data source) |
--model |
Set the model for this invocation without /model |
--permission-mode |
Start the session in a named mode (plan, dontAsk, …) |
--allowedTools |
Pre-approve specific tools or command patterns for the run |
--cloud |
Start a cloud session on Claude Code on the web (clones the repo’s current branch) |
| Shortcut |
What it does |
Shift+Tab |
Cycle the modes available in the current configuration, normally default → acceptEdits → plan |
Escape |
Interrupt the agent mid-response; Esc Esc with an empty prompt opens the rewind menu |
Ctrl+C |
Cancel the current input or action |
| Mode |
Behaviour |
default |
Reads are free; edits and most shell commands ask for approval |
acceptEdits |
File edits and common filesystem shell commands auto-approve; other shell commands still ask |
plan |
Reads run freely; commands prompt you or are gated by the safety classifier, and approved commands can run; edits are held until you approve a plan |
auto |
Most actions auto-approve; a background classifier blocks the unsafe ones (documented for current versions; see the permission-modes reference) |
dontAsk |
Deny calls that would otherwise prompt; useful for unattended runs |
bypassPermissions |
Skips most permission prompts; use only inside strong isolation |
A plain claude launch in the terminal or VS Code extension starts in auto mode (v2.1.283+; earlier versions did this only on Pro, Max, and Team plans); --permission-mode pins a different starting mode, and claude -p and SDK runs start in default unless you set one. The permission-modes reference lists the exceptions. Deny rules block in every mode, including bypassPermissions.
| File |
Purpose |
CLAUDE.md |
Project instructions loaded at session start (rules, conventions, context) |
.claude/settings.json |
Per-project permission rules, allowed/denied tools and paths |
.claude/skills/ |
Reusable skill definitions the agent invokes by description match |
.claude/agents/ |
Custom subagent definitions with their own model and permissions |
