You pay for Claude Code every month. Then you sit in front of it and approve every single step — write this file, okay, now run the test, okay, now fix the error, okay. That is the slow lane. The Claude Code slash commands that matter most do the opposite: they let the AI plan, work, and check its own code for hours while you go do something else. Most people never touch them.
Quick Answer
The Claude Code slash-command stack is four native commands — /ultraplan, /goal, /agents, and /ultrareview — that turn Claude Code from a tool you prompt step by step into an agent that runs itself. /ultraplan drafts a full plan, /goal loops toward a finish line on its own for hours or days, /agents pushes that job to the background so you can start a second one, and /ultrareview checks all the code the loop produced. All four are included in the paid plan you already have. The shift is from chatting with the AI to setting a goal and walking away — then coming back to finished, reviewed work. Anthropic reports 80% of its own new production code is now written by Claude, so this is not a trick; it is how serious work already ships.
You're babysitting an AI that's built to run on its own
Here is the pattern almost everyone falls into. You open Claude Code, give it one instruction, watch it work, approve the result, then give it the next instruction. You are the loop. The AI does a burst of work and then waits for you, over and over, for the entire session.
It feels productive because you are busy. But you are spending your attention on something the tool can do for itself. You are the bottleneck standing between a goal and a finished result — and you are paying for the privilege of being that bottleneck.
Why this is the expensive way to use Claude Code
The gap is not about who codes faster. It is about who lets the agent run.
Definition
An autonomous agent loop is when an AI keeps working toward a goal on its own — drafting, executing, checking, and correcting — turn after turn, without a human approving each step. Claude Code's /goal command can hold that loop for hours, and in long jobs, for days.
The proof is not hypothetical. Anthropic reports that 80% of its own new production code is now authored by Claude. Claude Code now writes roughly 4% of all public commits on GitHub, with daily commit volume up about 200% in the eight weeks ending mid-May 2026. One developer study clocked a 12x speedup on a task — 14.8 minutes with the agent versus 3.8 hours without.
Read those numbers again. The people producing that volume are not typing faster. They set a goal, let an agent loop on it, and review the output at the end. If you are still approving every step by hand, you are competing against people who stopped doing that months ago.
The four commands that turn Claude Code into an operator
There is a small stack of Claude Code slash commands that move you from "prompter" to "operator." You do not need all of Claude Code's surface area. You need these four, and you need to understand what each one does.
| Command | What it does | Runs where |
|---|---|---|
/ultraplan | Drafts the whole plan before any code is written | Cloud swarm (local-recreatable) |
/goal | Loops toward a finish line on its own, for hours or days | Your session |
/agents | Backgrounds the running loop so you can start a second one | Your session |
/ultrareview | Reviews all the code the loop just produced | Cloud swarm (local-recreatable) |
/ultraplan — it drafts the whole plan for you
Instead of you breaking a big job into steps, /ultraplan sends a swarm of helper agents to draft the entire plan first. You start from a structured plan, not a blank prompt. (Version 2.1.91 and up.)
/goal — it works toward the finish line on its own, for days
This is the heart of the stack. You give /goal a finish line, and Claude keeps working toward it on its own, turn after turn. It does not stop and wait after every action. On a real job it can run for hours; on a big one, for days, until the goal is actually met. (Version 2.1.139.)
/agents — it runs the job in the background while you start another
Once a loop is running, /agents pushes it into the background so you can open a second session right beside it. Now two jobs build at once instead of one job holding your whole screen hostage. Claude Code shipped an Agent View dashboard in May 2026 to watch these run side by side.
/ultrareview — it checks its own work when the loop is done
When the loop finishes, /ultrareview sends another swarm of agents to review all the code the loop just produced — catching mistakes before you ever see them. The agent that wrote the work is not the only agent that judges it. (Version 2.1.91 and up.)
Notice what is not on this page: the exact order you chain these in, and the one setup move that keeps your credit bill flat. That is deliberate. Knowing the four commands exist is the easy part. Wiring them into a single hands-off loop is the part worth getting right.
"Won't this cost a fortune in credits?"
This is the objection that stops most people, so let's be straight about it.
Claude Code is a paid plan. These four commands are already included in it — you are not buying anything new to use them. You are using capability you already pay for and currently leave on the table.
Two of the four — /ultraplan and /ultrareview — run as cloud swarms, which is where extra credit usage can creep in. There is a clean way to get the same planning and the same review on your own machine instead, so a long-running loop does not quietly run up your bill. That local setup is the difference between "this is amazing" and "this is amazing and it stays cheap." The exact recreation is in the free guide below — it is the single most practical part of the whole stack.
What it feels like to stop babysitting
Picture your next real job. You set the goal. You let the plan get drafted. You start the loop, push it to the background, and open a second session to work on something else. Hours later you come back, run the review, and read finished, checked work instead of approving fragments all afternoon.
That is the whole shift. You stop being the loop. The agent becomes the loop, and you become the person who decides what it builds.
Key Takeaways
- 4 native Claude Code slash commands — `/ultraplan`, `/goal`, `/agents`, `/ultrareview` — move you from prompting to operating.
- `/goal` can hold an autonomous loop for hours, and on large jobs, for days — no step-by-step approvals.
- 80% of Anthropic's own new production code is now written by Claude; Claude Code writes ~4% of all public GitHub commits.
- All 4 commands are included in the paid plan; 2 of them can be run locally so a long loop does not run up extra credits.
- The bottleneck is not the AI's speed — it is the human approving every single step.
Frequently Asked Questions
Are the Claude Code slash commands free?
No, and that distinction matters. Claude Code itself is a paid plan, so the engine is not free. The four commands are included in that plan at no extra charge, which means you are not buying an add-on to use them. Two of them, /ultraplan and /ultrareview, run as cloud swarms where extra credits can be spent. Both can be recreated locally to avoid that cost, so a long loop does not quietly run up your bill.
How long can /goal actually run on its own?
It runs until the finish line you set is actually met, not for a fixed block of time. For a small, focused task that might be a few minutes. For a substantial build it can hold the loop for hours, and on a large job it can keep going across days, working turn after turn without waiting for you to approve each step. The trick is writing a clear, checkable finish condition so it knows exactly when the work is done and when to stop.
Do I need to be an experienced engineer to use this? No. If you already use Claude Code to ship work, you have everything you need to run this stack. It does not add new coding requirements; it removes the manual approving you do now. The real hurdle is psychological rather than technical: trusting the loop enough to walk away while it runs. Watch your first two or three loops from start to finish, and that trust tends to build quickly once you see the output hold up.
Why review code the AI just wrote with another AI?
Because the agent that produced the work is not the best judge of whether it is correct. /ultrareview brings a fresh set of agents to inspect the output with no attachment to having written it, so they catch mistakes a single pass tends to miss. It is the same reason human engineering teams use separate reviewers instead of letting the author sign off on their own pull request. A second set of eyes, even synthetic ones, raises the floor on quality.
Read Next
- Claude Routines: Set It Once, Your AI Works While You Sleep — the scheduling side of hands-off Claude Code.
- What Are the Best AI Agents for Business Automation? — where autonomous loops fit in a wider automation stack.
- Claude Routines Replaced My Entire Automation Setup — a production look at running Claude unattended.
Next step
You now know the four commands and what each one does. What turns them into a machine that runs your work overnight is the exact order you chain them — and the local setup that keeps the credits in check.
Get the full 4-command set, the chaining workflow, and the credit-saving local recreation in the free guide: The Autonomous Claude Code Loop.