🏆 The Top 12

Set a goal with a "done" criterion, and Claude won't stop—while verifying—until it meets the condition.

You trigger it, go to sleep or grab lunch, and come back highly confident it’s ready. You don’t need to run it for hours.

Define "done" as objectively as possible; subjectivity is the Achilles' heel.

Offload planning to the cloud with multiple planning agents; review and comment in your browser.

Ideal for large, ambiguous tasks where planning before execution is worthwhile. Keeps working while it plans.

Runs in the cloud, brings the plan back to the terminal, or executes it directly on the web.

Generates an HTML report of your Claude Code usage over the last 30 days: what you did a lot, where you got stuck, and what to use more.

Shows patterns and friction points you don’t notice on your own—quick wins and ambitious workflows to try.

Check in from time to time to see how your usage evolves.

Previously called "autodream": Claude improves its memory on its own, without you asking, at intervals based on sessions/time.

It “dreams,” scans what’s been done, and gets smarter about each repository — without any work from you.

Check in /memory whether it’s enabled; turn it on if you want.

A named team of agents that run in parallel and talk to each other — panels, personas, debate boards.

Great for hearing different perspectives (beginner, CEO…) debate until they agree. Token-intensive.

Experimental feature: enable it in settings. Invoke it in natural language.

Roll back the code and conversation to an earlier checkpoint—you choose the exact point.

When something went wrong, rewinding is better than saying “no, do it this way”—it clears the polluted context.

Think of it as cached context: rewind preserves quality better than piling on corrections.

Background agents that run in parallel, each in its own context window; they report to the main terminal.

They do several things at once without cluttering your context. Unlike Agent Teams, they don’t talk to each other.

Definable as markdown in .claude/agents, with its own tools and skills.

Repeats a prompt at an interval until a condition is met, using the cron tool.

Lightweight recurring reminders and checks. No need for the /loop — you can ask in natural language.

Tied to the session (close it, and it’s gone) and turns off automatically after ~7 days.

Control your local Claude Code session from your phone or the web, synced in real time.

Go for a walk and keep the session alive; what you type echoes on both devices.

It’s the same session, not a copy. One link or the Claude app (Code tab).

Schedules a real agent (not a script) in cron — locally or in the cloud, with your laptop closed.

You get an agent that researches and makes decisions on a schedule, not just API calls.

/schedule creates; the desktop app manages them in a calendar and list.

Scriptable bottom bar: model, effort, and how much of the context window has been used.

The most underrated feature: you know when to hand off, /clear, or /compact.

/statusline and describe it (even with a screenshot). Combine it with /context and /usage.

Markdown recipes that teach Claude workflows. It opens the “recipe” and follows the steps.

The ecosystem’s biggest force multiplier: consistency > repeating prompts. Create your own and pull in other people’s.

Simple or complex, chainable, shareable—and they work in Code, Chat, and Cowork.