Week 13: Auto Mode, Conditional Hooks, and the First Big Update

New official middle ground between 'ask before every edit' and 'dangerously skip permissions'. Uses a risk-based classifier to auto-approve safe actions (reading files, running tests) while flagging risky ones (deleting files, running unknown scripts) for human review. Use this instead of --dangerously-skip-permissions for daily work.

claude config set autoMode true
# Or in settings.json:
"permissions": { "autoMode": true }

Press Ctrl+O to open the transcript picker, then type / to search across all past conversations. Finds old solutions, decisions, and code snippets from previous sessions. Use this when you remember discussing something but can't find which session it was in.

# In Claude Code terminal:
Ctrl+O    # Open transcript picker
/search   # Search across all conversations

Drag images into the chat and reference them with [Image #N] chips. Claude can analyze screenshots, mockups, and diagrams inline. Use this when you want Claude to match a visual design or debug a UI issue from a screenshot.

Press Ctrl+X then Ctrl+E to open your current prompt in your default $EDITOR (vim, nano, etc.). Edit complex multi-line prompts in a real editor, then save and close to send. Use this for long, detailed prompts that are hard to write inline.

# In Claude Code input:
Ctrl+X Ctrl+E    # Opens $EDITOR with current prompt
# Edit, save, close → prompt is sent

For enterprise teams: drop JSON files into /etc/claude-code/settings.d/ (Linux) or ~/Library/Application Support/ClaudeCode/managed-settings/ (macOS). Files merge alphabetically. Use this when you need to enforce team-wide settings that individual devs can't override.

# Linux: /etc/claude-code/settings.d/
# macOS: ~/Library/Application Support/ClaudeCode/managed-settings/

# Example: 01-security.json
{
  "permissions": {
    "deny": ["Bash(rm -rf *)","Bash(sudo *)"]
  }
}

Restrict which plugin channels (registries) Claude Code can install from. Prevents rogue plugins from untrusted sources. Use this in enterprise environments where you want to control which plugins your team can use.

// settings.json
{
  "plugins": {
    "allowedChannels": ["official", "your-org-channel"]
  }
}

Claude Code now has a native PowerShell tool for Windows users. Runs PowerShell commands directly instead of routing through WSL/Bash. Still in preview — use if you're on Windows and want native command execution.

Hooks can now use an `if` field with permission rule syntax to determine whether they should run. Drastically reduces unnecessary process spawning — hooks only fire when they match the specific tool pattern. Use this when you have hooks that should only trigger on specific commands (e.g., only intercept `rm` commands, not every Bash call).

"PreToolUse": [
  {
    "matcher": "Bash",
    "if": "Bash(rm *)",
    "hooks": [{"type": "command", "command": "python3 ~/.claude/hooks/safety_check.py"}]
  }
]

Three new hook events for reactive environment management. TaskCreated fires when a new task starts (setup environment). CwdChanged fires when the working directory changes (load project-specific configs). FileChanged fires when Claude modifies a file (auto-lint or auto-format). Use these when you want hooks that respond to environment changes, not just tool calls.

"TaskCreated": [{ "hooks": [{"type": "command", "command": "echo Starting task..."}] }],
"CwdChanged": [{ "hooks": [{"type": "command", "command": "cat .claude/project-rules.md"}] }],
"FileChanged": [{ "matcher": "*.ts", "hooks": [{"type": "command", "command": "npx eslint --fix $FILE"}] }]

A PreToolUse hook that intercepts any file deletion or destructive command and requires explicit confirmation. Prevents Claude from accidentally deleting important files. Use this on any project where accidental deletion would be costly.

"PreToolUse": [
  {
    "matcher": "Bash",
    "if": "Bash(rm *) || Bash(git reset --hard *)",
    "hooks": [{
      "type": "command",
      "command": "echo 'BLOCKED: Destructive command detected. Please confirm.' && exit 2"
    }]
  }
]

After 75+ minutes of inactivity, Claude Code now shows a prompt asking if you want to continue the current session or start fresh. This prevents accidentally burning tokens on a stale context. Just be aware it exists — no setup needed.

Create a team.md file that defines agent roles, responsibilities, and handoff protocols. Each agent reads team.md to understand who does what. Use this when running 3+ agents that need to coordinate without stepping on each other's work.

# team.md
## Agent Roles
- **architect**: Owns system design, creates PRDs
- **frontend**: Implements UI components, styling
- **backend**: API routes, database, auth
- **reviewer**: Code review, testing, quality

## Handoff Protocol
1. Architect creates PRD → drops in docs/
2. Frontend/Backend read PRD → implement
3. Reviewer validates → creates issues
4. Use git branches per agent to avoid conflicts

Use tmux split panes to monitor multiple Claude Code agents simultaneously. Each pane runs a different agent on a different worktree. Use this when running 2-4 agents in parallel and you want to see what each one is doing in real-time.

# Create tmux session with 4 panes:
tmux new-session -s agents
tmux split-window -h
tmux split-window -v
tmux select-pane -t 0 && tmux split-window -v

# In each pane, cd to a different worktree:
# Pane 0: cd project-main && claude
# Pane 1: cd project-frontend && claude
# Pane 2: cd project-backend && claude
# Pane 3: cd project-tests && claude

A progression framework: L1 (plain prompt), L2 (+ reference images), L3 (+ component library), L4 (+ design system), L5 (+ Figma MCP), L6 (+ Stitch design bridge), L7 (+ custom skill + hooks for auto-formatting). Each level produces significantly better output. Use this to understand where you are and what to add next.

Chrome extension that lets you draw directly on your localhost preview and send the annotated screenshot to Claude. Circle a button, draw an arrow, write 'make this bigger' — Claude sees your visual feedback. Use this when text descriptions of UI changes aren't precise enough.

Record a Loom video of your workflow, feed it to Gemini to extract step-by-step instructions, then convert those instructions into a Claude Code skill. Use this when you have a complex manual workflow that you want to automate but it's hard to describe in text.

The most-starred Claude Code community repo. Comprehensive collection of best practices, CLAUDE.md templates, and workflow patterns curated by the community. Use this as a reference when setting up a new project.

Collection of 135+ pre-built agent configurations for different use cases (frontend, backend, DevOps, data science, etc.). Use this when you want a ready-made agent config instead of building one from scratch.