Is Loop a Claude Code-exclusive feature? Can regular Claude.ai users not access it?
Currently, the Loop functionality described in Anthropic's official guide — including the /goal, /loop, and /schedule commands — is primarily available in Claude Code, the standalone command-line tool. If you only use Claude.ai's web chat, these commands aren't directly accessible.
But this doesn't mean the Loop concept is irrelevant to you. Claude Cowork (Claude's workflow automation tool) and Claude Tag (Slack integration) follow the same design direction described in the Loop guide — all of them let Claude pursue goals autonomously rather than waiting for your instruction at every step. If you use those tools, you're already using an application of Loop concepts.
For regular Claude.ai users, the most practically relevant takeaway: in Claude Projects, write the definition of "how to know the task is complete" into your Project Instructions. This is the simplest way to give Claude self-checking capability without an automatic Loop mechanism.
How do I set a reasonable 'maximum attempt count' for Goal-based mode?
There's no fixed answer, but a few frameworks help you decide.
First, consider "how much does one attempt cost?" If each round requires running a full test suite, deploying to a test environment, and running performance analysis, one round is expensive — setting 3-5 is reasonable. If each round is just modifying a few lines of code and running a simple check, 10-15 is fine.
Second, consider "if the goal isn't reached, do I want to know or have it keep trying?" When Goal-based hits the limit, it doesn't silently give up — Claude tells you "tried X times and still haven't reached the goal," letting you decide whether to adjust the target, continue with more attempts, or change direction. Setting a conservative limit and seeing where 5 attempts get you is smarter than setting 20 upfront — you get feedback faster on whether you're going the right direction.
Third, the difficulty of the goal affects what's reasonable. Getting a Lighthouse score from 75 to 90 might work in 5 attempts; raising test coverage from 60% to 80% might need more rounds. Before you have a feel for the task's difficulty, start conservative.
How do I track token usage and costs? Can I stop a Loop midway?
Yes, you can stop it, and Anthropic includes specific tracking tools in the guide.
Tracking tools:
/usage command: shows recent token usage including which skills were used, whether subagents were dispatched, and which MCP tools were called./goal (with no parameters): shows how many rounds the current goal has run and how many tokens have been consumed./workflows: shows each agent's token usage and lets you stop specific agents from this interface.Stopping midway: Yes — anytime you feel things have gone off-track or costs are too high, you can interrupt. Claude Code has a kill mechanism, and the /workflows interface also lets you stop running agents.
Practical cost control: The guide specifically emphasizes that the right mindset for Loop cost management is "pilot small, confirm direction is right, then scale." Especially with Proactive mode (event-triggered, multi-agent parallel) — token consumption can be very fast. Before you're familiar with the mode's behavior, don't deploy at full scale on a large project. Start with a small pilot, observe per-round consumption, and then decide whether to expand.
Do these four Loop modes have any relevance for non-developer knowledge workers?
Yes — more directly than you might expect. The specific commands in this guide (/goal, /loop, etc.) are Claude Code features, but the design thinking behind them is useful for anyone who wants to use Claude more effectively.
Insight 1: Clarify 'how to know it's done.' The core recommendation of Turn-based mode is: don't just tell Claude what to do — tell it how to confirm whether it did it right. This applies to any Claude use case. When you give Claude a task, adding "if done correctly, the output should meet these conditions..." improves Claude's self-verification accuracy.
Insight 2: Give a goal, not a list of instructions. Goal-based mode has you say "reach this result" rather than "follow these steps." This thinking applies in regular Claude conversations too — describing the result you want usually produces better output than describing the process you want to follow.
Insight 3: Consider automating repetitive work with Cowork or Claude Tag. If you have tasks you repeat every week, this guide tells you that automating them is possible — just the entry point isn't Claude Code, it's Cowork or Claude Tag in Slack.
Finally, the biggest takeaway from this guide: Claude's ultimate capability ceiling is significantly higher than what you see from everyday chat interactions. Understanding that ceiling exists helps you think bigger — not just 'can Claude help me do this?' but 'if I design it well, what can Claude accomplish autonomously within what scope?'
On July 6, 2026, Anthropic's official developer account ClaudeDevs posted a technical guide on X about using Loops to get more out of Claude Code. Written by Claude Code team engineer Delba de Oliveira, it takes months of developer community experimentation and organizes it into four distinct patterns.
It sounds technical, but the core concept is simple: today, every time you use Claude Code, you say one thing, it does one thing, you say the next thing. With Loops, once you give Claude Code a goal or a condition, it repeats, checks, and corrects on its own until the task is genuinely done — without you hovering over it the whole time.
What does this mean for non-engineers? Imagine telling an employee "improve this page's performance score to 90 or above." You don't need to spell out every step — they figure it out, test it, keep going if it's not there yet, and report back when it's done. Loops are what let Claude Code work that way.
Loops aren't a new concept technically, but Anthropic building them into Claude Code in an accessible way is what this guide is really about.
In developer circles, Loops started being widely discussed around late 2025 — more and more people were discovering that letting Claude run toward a goal automatically produced dramatically better results than issuing one manual command at a time. This guide is Anthropic taking that community knowledge and turning it into official documentation for a wider audience.
The four modes run from "requires the most manual intervention" to "most automated": Turn-based (manual loop), Goal-based (objective-driven), Time-based (scheduled execution), and Proactive (event-triggered).
This is essentially how you already use Claude Code — just with an explicit verification step added in.
You give Claude a task, it executes, and then it confirms whether it did it correctly, rather than just handing the result back to you. For example, you ask it to modify a frontend page — instead of just making the change and delivering it, it starts a dev server, checks the browser for issues, confirms the console has no new errors, and only tells you it's done once everything checks out.
The guide's recommendation: write down "how to know whether you got it right" in your SKILL.md file. This way Claude always knows what checklist to self-verify against when it finishes a task, without you needing to re-explain every time.
For non-developers, the concept is this: tell Claude the result you want, and also tell it how to know whether it got there. Many people skip this second part, which is why Claude finishes a task and you're still not satisfied — leaving both sides confused.
This mode is triggered by the /goal command. You tell Claude a specific target and a maximum number of attempts, and it keeps trying until it hits the target or exhausts its attempts.
The guide's example: /goal get the homepage Lighthouse score to 90 or above, stop after 5 tries. After each attempt, Claude checks the score — still below 90, it keeps modifying; hits 90, it stops.
There's an interesting design detail: when Claude wants to stop, a second "evaluator model" confirms whether the goal was actually met — a second AI watching to prevent Claude from prematurely concluding it's done when it isn't.
This mode works best for tasks with clear numeric targets: test pass rates, coverage thresholds, performance scores. The more specific the goal, the more efficient the mode — and the less compute it burns.
This turns Claude Code into a scheduled automation assistant. You use the /loop command to tell it how often to do something and what to do.
The guide's example: /loop 5m check my PR, address review comments, and fix failing CI. Every 5 minutes: automatically check your pull request, address any review comments, fix failing automated tests. You go to a meeting; it keeps monitoring and handling things in the background.
Running on your own machine, /loop is good for temporary tasks — close your computer and it stops. If you want the task to keep running on the cloud even with your computer off, use /schedule instead to create a persistent routine. Good for: automatically summarizing yesterday's notifications every morning, hourly service monitoring, and similar recurring work.
This is the most powerful — and the most careful-to-configure — of the four modes. Instead of being triggered by time, it's triggered by something happening.
The guide's compound example: set a routine that automatically scans a user feedback channel every hour; whenever it finds a bug report, use /goal to define what "fully handled" means for that bug; then automatically dispatch multiple AI Agents in parallel to investigate, fix, and report on different bugs simultaneously. The entire flow from "user reports a bug" to "bug triaged and addressed" runs without manual intervention at each step.
This sounds compelling — and it is — but the guide specifically emphasizes: this mode runs many AIs simultaneously and compute costs are significant. The strong recommendation is to pilot in a limited scope first, confirm the process design is working, and only then expand to a full project.
If you currently only use Claude.ai's standard chat features, this Loop guide is primarily about Claude Code (a standalone developer tool) and doesn't directly apply to your current context. But the direction it points to is worth understanding.
The bigger picture: this guide describes AI evolving from "you prompt it once, it does one thing" toward "give it a goal and let it figure out how to get there." This direction shows up not just in Claude Code but in Claude Cowork, Claude Tag, and other tools — AI's role shifting from "tool that waits for your commands" to "collaborator that completes tasks autonomously."
If you're a developer or interested in going deeper with Claude Code, this guide is the most complete official documentation available right now. If you're not a developer, the takeaway is simpler: the range of tasks Claude can automate is expanding fast. What Claude can do for you today is only a subset of what it'll be able to do in the near future.