Level 3 Is Three Different Schedulers, And Only One Runs While You Sleep
Most “three levels” cards sort by how advanced you are. This one does something better: it sorts by who starts the thing and where it runs. Level 1 is a skill you fire from a chat. Level 2 is a timer, firing “with your laptop open”. Level 3 is “on the cloud with your laptop closed”.
That is the right axis, and the subtitle states it cleanly: “Level 1 runs from one command. Level 3 runs while you sleep.” It is also a claim you can check, because Anthropic publishes a comparison table of its scheduling options. Checking it turns up something the card’s single arrow cannot show: Level 3 is not one thing. It is three, and two of them still need your machine switched on.
Source card: “The 3 Levels of Claude Automation” — BANI Academy, 18 October 2026. Level 1: Skills in Claude Chat. Level 2: Scheduled tasks in Cowork. Level 3: Claude Code, with /schedule.
View the original post on Facebook
Diagram: the three schedulers hiding inside Level 3
The Claude Code documentation opens its scheduling page with a line that settles the question: “Claude Code offers three ways to schedule recurring or one-off work.” Here are the rows that decide which one you want.
| Routines (cloud) | Desktop scheduled tasks | /loop |
|
|---|---|---|---|
| Runs on | Cloud, Anthropic-managed by default | Your machine | Your machine |
| Requires machine on | No | Yes | Yes |
| Requires open session | No | No | Yes |
| Access to local files | No (fresh clone) | Yes | Yes |
| Minimum interval | 1 hour | 1 minute | 1 minute |
| Permission prompts | No — runs autonomously | Configurable per task | Inherits from session |
| How you set the schedule | Via /schedule in the CLI |
Yes | Yes |
Credit where it is due: the card picked the right one. /schedule does create a cloud routine, it does run with the laptop shut, and “manage it from the browser: pause, edit or check every run from anywhere” matches what routines are for. The card is not describing a feature that does not exist.
What it does not say is what that column costs.
The one that runs with the laptop closed cannot see your files
Look at the fourth row again. Cloud routines have “Access to local files: No (fresh clone)”. The thing that runs while you sleep starts from a clean checkout. It does not have your desktop, your notes folder, your half-finished spreadsheet or the vault you spent a weekend building.
That matters because of how the card frames Level 3: “Everything else, runs while you sleep. Any tool, any API, any hour on the cloud.” Read as a ladder, Level 3 sounds like a superset — everything Level 2 did, plus the laptop can be shut. It is not a superset. It is a trade.
| What you want to automate | Which of the three | Why |
|---|---|---|
| Something that touches files on your machine — a folder, a repo, a local vault | Desktop scheduled task | It is the only option with both local file access and no requirement for an open session. The cloud gets a fresh clone instead |
| Something that must happen whether or not the laptop is open — a daily report from APIs and connectors | Routine (cloud) | The one column with “Requires machine on: No”. Note the 1-hour minimum interval |
| Something more often than hourly | Not the cloud | 1 hour is the floor there. The two local options go down to 1 minute |
| Watching a build or a PR for the next half hour | /loop |
Cheapest to start, needs the session open, and it is designed for exactly this |
| Anything you want still running next month | Routine or Desktop task | See the next section. /loop has a fuse on it |
If you reach for /loop first, know about the seven days
/loop is the one a beginner will find first, because it is a single command typed into a session you already have open. Two documented behaviours are worth knowing before you rely on it.
“Recurring tasks automatically expire 7 days after creation. The task fires one final time, then deletes itself. This bounds how long a forgotten loop can run.”
“Tasks only fire while Claude Code is running and idle. Closing the terminal or letting the session exit stops them firing.”
Both are sensible designs — the expiry exists so a loop you forgot cannot bill you forever. But together they mean /loop is a session tool, not an automation. If what you wanted was “runs while I sleep”, this is not the one, and it will look like it is working right up until the terminal closes.
There is a third behaviour worth a line, because it surprises people: scheduled tasks are deliberately jittered. The docs state that a recurring task “starts each run late by up to half the gap to the next run, and by at most 30 minutes”, to stop every session on earth calling the API at the same second. An hourly task is not an on-the-hour task.
The two claims on the way up the ladder
Two smaller things on the card are worth checking, because they are the sentences that make the ladder feel seamless.
| The card says | What to know | Verdict |
|---|---|---|
| Level 1: “Claude writes the whole skill, you never touch the file” | True in the sense that you can describe a workflow and have the skill authored for you. It is still a file, and whether you can edit it later depends on which surface you built it in | Fair, with the note that “never touch the file” is a convenience, not a guarantee you will never need to |
| Level 3: “one terminal command, your Level 1 skills carry straight over” | Skills are a shared format, so the idea is sound. But a chat-authored skill and a Claude Code skill live in different places, and a cloud routine starts from a fresh clone — so “carry over” means you supply them, not that they appear | Check it yourself before planning around it |
| Level 2 is named “Scheduled tasks in Cowork” | The release notes for 16 September say Cowork is merging into regular chat, with tasks, projects, connectors and skills carrying over. We went through this in an earlier piece | The capability stays; the name on the card may not |
And the card’s own footer anticipates all of it, in grey type at the bottom: “Features, commands and plan limits (Cowork, /schedule, Routines) change often. Check Claude’s official docs for what is available to you.” It even names Routines — the correct term for the cloud column — which is more precision than the body of the card uses.
Four honest caveats
- We did not run any of the three. Everything above is the card checked against the published comparison table and the scheduled-tasks page, in October 2026.
- “Cloud, Anthropic-managed by default” has a default in it. Self-hosted environments exist, and a team running its own runners would get different answers on the local-files row. The table describes the standard setup.
- The 1-hour minimum applies to the cloud option as documented. Minimums are exactly the kind of number that gets relaxed quietly, so check before concluding you cannot have what you need.
- The card is not selling a fantasy.
/schedule, cloud routines and browser management are all real and all work as described. Our objection is to the arrow, not the destination: three mechanisms drawn as one rung hides the trade-off that decides which you should pick.
The one-line version
The card’s axis is right — on demand, then on a timer, then while you sleep — but the top rung is three different schedulers, and the only one that truly runs with your laptop closed is also the only one that cannot see a single file on it, and cannot run more than once an hour.
Pick by two questions, not by level: does this job need my files, and does my machine need to be off? If the answer to both is yes, no rung on this ladder has it.
Source card: “The 3 Levels of Claude Automation” — BANI Academy, 18 October 2026.
View the original post on Facebook · Related: another card whose last step rested on a preview feature




One Comment