The features that shipped in the last 6 months -- Plan Mode, session resume, scheduled tasks, model switching, and the web app. Updated monthly.
Claude Code ships fast. Features that didn't exist 6 months ago are now core to how power users work.
This lesson is updated monthly. If a feature you heard about isn't here, it may have shipped after this was written -- check the official changelog at code.claude.com/docs/en/changelog.
What it is: A read-only analysis mode. Claude can read your files and think through a plan, but cannot edit anything until you approve.
Why it matters: Before Plan Mode, Claude would just start making changes. Now you can see exactly what it intends to do first -- and say no to any step.
How to use it:
| Method | How |
|---|---|
| Cycle modes mid-session | Press Shift+Tab (Normal -> Auto-Accept -> Plan Mode) |
| Start a session in Plan Mode | claude --permission-mode plan |
| Set as project default | Add defaultMode=plan to .claude/settings.json |
| In VS Code | Plan opens as a full markdown doc you can annotate |
Best use cases: Multi-file refactors, exploring an unfamiliar codebase, any time you want Claude to think before it touches anything.
The old problem: Every new terminal session started fresh. No memory of what you were doing.
The fix:
In practice: You close your laptop. Next morning, type claude --continue and pick up exactly where you stopped.
What it is: Manually trigger context compaction -- Claude summarizes the conversation to free up context window space.
Type: /compact in any active session.
Claude also compacts automatically when needed. A PostCompact hook event fires after compaction, which is useful for injecting fresh context into a long-running session.
Three ways to run Claude on a schedule:
1. Cloud scheduled tasks (most reliable)
Run on Anthropic infrastructure even when your computer is off.
Configure at claude.ai/code or via /schedule in the CLI.
Best for: morning PR reviews, overnight audits.
2. Desktop scheduled tasks
Run on your own machine via the Desktop App.
Best for: tasks that need access to local files.
3. /loop (session-bound)
Repeats a prompt at fixed intervals within the current CLI session.
Example: /loop 5m check if any tests are failing
Maximum 3-day expiry. Cancelled when you close the terminal.
You can now switch models mid-session without restarting.
| Model | Best for |
|---|---|
| claude-sonnet-4-6 | Default -- handles 90%+ of coding tasks fast and cost-effectively |
| claude-opus-4-6 | Deep reasoning, complex architecture decisions, 1M token context |
| claude-haiku-4-5 | Quick checks, simple reads, budget tasks |
How to switch:
- Mid-session: type /model and pick from the list
- One session: claude --model claude-opus-4-6
- Permanent: export ANTHROPIC_MODEL=claude-sonnet-4-6 in your shell config
- VS Code: Extension settings -> selectedModel, or /model in the chat panel
Effort levels (Sonnet and Opus only):
Type /effort to set low / medium / high. Type ultrathink in any prompt for maximum reasoning on that single turn.
Multiple Claude Code instances working in parallel on different parts of a task.
The command syntax changed. Use these formats:
# STDIO transport (local server -- most common)
claude mcp add --transport stdio <name> -- <command> [args...]
# HTTP transport (remote server)
claude mcp add --transport http <name> <url>
# Scope options
--scope user # available in all your projects (default)
--scope project # saved to .mcp.json, shared with teammates
--scope local # only you, only this repo
On macOS (Pro or Max subscription, Desktop App required), Claude can:
- Open apps and click buttons on your screen
- Control Chrome: click, read console logs, fill forms, automate across tabs
Not available on Windows or Linux.
Official changelog: code.claude.com/docs/en/changelog -- updated with every release.
HivePowered AI — AI Like a Pro Training