Latest capabilities and how to use them with pro-workflow.
Talk to Claude Code instead of typing. Rolling out to Pro, Max, Team, and Enterprise plans.
/voice # Toggle voice mode on/off- Hold spacebar to talk, release to transcribe
- Mix typed and spoken input in a single prompt
- Works alongside all other features (hooks still fire, modes still apply)
| Scenario | Why Voice |
|---|---|
| Describing architecture | Faster than typing long descriptions |
| Quick fixes | "Fix the import on line 42" |
| Exploring ideas | Stream-of-consciousness brainstorming |
| Reviewing code | "Walk me through this function" |
Voice mode works with all pro-workflow patterns. Corrections spoken verbally still trigger the self-correction loop if you say "remember that" or "add to rules".
Coordinate multiple Claude Code sessions working as a team.
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}- One session is the lead — coordinates work, assigns tasks
- Teammates work independently in their own context windows
- Teammates message each other directly (not just report to lead)
- Shared task list with dependency tracking
| Mode | How | Best For |
|---|---|---|
| In-process | Shift+Down to navigate |
Quick coordination |
| Split panes | tmux or iTerm2 splits | Visual monitoring |
Press Shift+Tab to toggle delegate mode. Lead orchestrates only, no direct code edits.
| Pattern | Lead | Teammate A | Teammate B |
|---|---|---|---|
| Frontend/Backend | Coordinate | UI changes | API changes |
| Feature/Tests | Plan | Implementation | Test coverage |
| Competing hypotheses | Evaluate | Approach A | Approach B |
| Cross-layer | Integrate | Database | Application |
https://code.claude.com/docs/agent-teams
Claude Code automatically tracks checkpoints during your session.
Esc Esc— Rewind to the last checkpoint/rewind— Browse and select a checkpoint to restore- Checkpoints are git-based — your code state is saved at each point
- Claude went down the wrong path
- An approach isn't working out
- You want to try a different solution
- Something broke that was previously working
The PreCompact hook saves context state, which complements checkpointing. Between checkpoints (code state) and pre-compact saves (context state), you can recover from most mistakes.
Continue sessions from phone, tablet, or browser.
- Start Claude Code on your machine
- Access it remotely via
--remoteflag - Continue interacting from any device
- Session state persists
Run Claude Code without a terminal UI for CI/CD and automation:
claude --print "review and fix lint errors" --max-turns 10Claude Code has expanded from 8 to 18+ hook event types.
| Event | When | Use Case | In hooks.json |
|---|---|---|---|
SubagentStart |
Subagent spawns | Log agent activity, set up context | Yes |
SubagentStop |
Subagent finishes | Collect results, update metrics | Yes |
TaskCompleted |
Task marked complete | Quality gate on completion | Yes |
PermissionRequest |
Permission dialog | Flag dangerous operations | Yes |
PostToolUseFailure |
Tool use fails | Error tracking, retry logic | Yes |
TeammateIdle |
Team member goes idle | Force continuation, reassign | Yes |
ConfigChange |
Settings modified | Detect mid-session changes | Yes |
Setup |
Initial setup (30s timeout) | One-time initialization | No |
WorktreeCreate |
Worktree created | Set up worktree-specific config | No |
WorktreeRemove |
Worktree removed | Cleanup | No |
Events marked "No" are available in Claude Code but not configured in hooks.json by default. Add them to your hooks.json or settings.local.json as needed.
| Type | How It Works |
|---|---|
command |
Run shell script, receive JSON on stdin |
prompt |
Single-turn model evaluation (Haiku default) |
agent |
Multi-turn subagent with Read/Grep/Glob (up to 50 turns) |
{
"PostToolUse": [{
"matcher": "tool == 'Edit'",
"hooks": [{
"type": "agent",
"agent": "reviewer",
"prompt": "Review this edit for security issues",
"timeout": 30000
}]
}]
}{
"PreToolUse": [{
"matcher": "tool == 'Bash' && tool_input.command matches 'rm'",
"hooks": [{
"type": "prompt",
"prompt": "Is this rm command safe? Check for wildcards and critical paths.",
"model": "haiku"
}]
}]
}| Command | Shows |
|---|---|
/usage |
Plan limits and remaining |
/extra-usage |
Pay-as-you-go overflow ($2,000/day cap) |
/cost |
Current session cost (API key users) |
Fast mode is always billed to extra usage from the first token, even if you have remaining plan usage. Toggle with /fast.
claude --max-budget-usd 5.00 # Cap at $5
claude --max-turns 50 # Limit turns