Monday, 6:40am. First dog at eight. The lights are half on, the dryer is cold, and for the next hour and twenty minutes there is nobody in the building to talk to.
Twenty of those minutes are genuinely quiet, and here is what usually happens to them. Coffee. Phone. The weekend's messages, two of them half answered. A look at the week ahead, a vague sense of being behind, and a decision that the numbers will get done properly on Wednesday. Wednesday already has a dog on it.
You can automate recurring tasks with AI, but not this. The twenty minutes are not the problem. What they get spent on is. They get spent deciding.
The direct answer, if you want to automate recurring tasks with AI. Write the routine down first, as a fixed list of jobs in a fixed order, so that Monday stops being a decision at all. Then automate only what honestly can be. Claude Code's first-party automation is a short list: Routines, reached by typing /schedule, on a paid Claude subscription; /loop with an explicit interval, which every user has; hooks, which fire on Claude Code's own lifecycle rather than on a clock; and GitHub Actions or GitLab CI if you already run a pipeline. Everything past that list is you wiring up your own computer's timers, which is legitimate and a bigger job than it sounds. And for most readers the right answer is none of the above: you open it at 6:40am and type one thing.
That last one is not a failure. It is most of the win.
This post assumes you know what Claude Code actually is, that you have got it installed, and that you have one skill of your own or borrowed. Our examples follow Steph, a two-groomer salon in Sydney's Inner West, a worked example we follow through this series rather than a real client's books. The 6:40am above is hers, and so are the numbers at the end.
Automate the decision before you automate the doing
Two different things get called automation, and only one of them is hard.
The first is execution without you: a computer runs the job at 6:40am while you are still in the car. The second is deciding without you: the order is already settled, so you never stand in the salon choosing.
The second is worth more, and it costs nothing but a piece of paper. A written routine is a decision you made once, on a calm Sunday, that keeps applying on a chaotic Monday. It survives a bad night's sleep, Moose arriving matted at 8:15, and a phone that will not stop.
It is the same argument as a CLAUDE.md, and it is Anthropic's own stated reason for writing a skill: you write one when you notice you are re-pasting the same instructions into a chat over and over.
One naming note. Anthropic calls its scheduling feature Routines, capital R. Your Monday routine is a different animal, and it comes first.
Build the routine: two jobs on a Monday, three on the first Monday of the month
Here is a real sequence, in the order that works.
- Last week's numbers. What you took, how many dogs came through, how many did not, and the one figure that moved most against the week before.
- Who is overdue for a groom. Worked out from each dog's own cycle rather than a blanket six weeks, with one draft message per client.
- What is unpaid. This one is monthly, not weekly. It goes on the first Monday of the month and nowhere else.
The order is not arbitrary. The numbers come first because they are the only job that changes what you do next. The overdue list comes second because it produces work you can act on the same day. Money owed is third because chasing it weekly is how you turn a client into an enemy over $95.
This is what you type at 6:40am.
On the first Monday of the month, one more line.
In this series those three jobs are three named skills: /weekly-numbers, /rebook-gaps, and /month-end. Run your own if you have written one. If not, the twelve-skill library is where to start, and how to build one is its own post.
What is genuinely first-party when you automate recurring tasks with AI
This is where most articles get loose. Here is the whole list from Anthropic's own documentation, with what each one means for a salon.
Routines, reached by typing /schedule. The first-party feature for running Claude Code on a schedule.
It is available on a paid Claude subscription. If you authenticate with an API key from the Anthropic Console instead, the command is not there at all, and the documented alternative is /loop. Which plan you are on is post 8's territory; the point is that your login decides whether this option exists.
/loop, with an explicit interval. Every Claude Code user has this one. It runs a prompt or a command repeatedly, written the way the documentation writes it: /loop every 2 hours. On a paid subscription, /loop on its own falls back to a default maintenance prompt and picks its own interval; on an API key or a cloud provider it shows you the usage message, with no default interval available.
Read the word interval again. The documentation describes an interval, not a day of the week, so if Monday specifically is what you want, /loop is not the thing that gives it to you.
Hooks. Shell commands Claude Code runs at fixed points in its own lifecycle: before a file edit, after a tool runs, when it is waiting on you, and other specific events. Their value is that they are deterministic. They run regardless of what the model decided to do, which is exactly the property a CLAUDE.md does not have, because that file is advice.
Hooks are not a clock, though, and a lot of writing about automation blurs this. A hook cannot make Monday happen; it can only make something happen once Monday already has.
GitHub Actions and GitLab CI. Genuinely first-party integrations for running Claude Code in the cloud. Also genuinely developer infrastructure. If you already run a pipeline you know what this is for. If you do not, this is not the Monday to find out.
That is the list. Everything else is you.
What you would have to assemble yourself to automate recurring tasks
Anthropic's own position on this is refreshingly blunt: automation beyond /schedule, /loop, and hooks means setting up your own infrastructure, and scheduling on unattended infrastructure is your responsibility.
A scheduled task on Windows or a cron job on a Mac is not a Claude Code feature. It is your operating system's timer, pointed at a command. It works, and the moment you build it you own it. Nothing in anyone's documentation will help you the week it quietly stops firing and you do not notice for a fortnight.
There is a more mundane version of the same problem. A scheduled job needs a machine that is awake and online, and Anthropic lists an internet connection among the requirements. A laptop shut in a bag has neither.
None of that is an argument against building it. It is an argument for being clear that you are building something, rather than switching something on.
The third answer, which is the one most of us should pick
You open it at 6:40am and you type one thing.
That is the honest answer for most readers, and it is worth defending rather than apologising for. The routine already made the decision, which was the expensive part. The twenty minutes now go on the jobs instead of the choosing. And the output gets read by a person the second it appears, which is a property an unattended routine has to work quite hard to fake.
If you want one more notch of automation without owning any infrastructure, the move is not a scheduler. It is to make the routine itself a skill, so the one thing you type is one word and the order lives in a file rather than in your memory of this article.
Nothing that sends, charges, or posts goes in an unattended routine
Ever. Two lines from the never list, which post 4 owns and nobody in this series reworded:
And never let it:
Sends, charges, posts, deletes. Four verbs, and not one of them has an undo. A routine running on a timer is precisely the arrangement where a wrong one goes out to everyone on the list before anybody is awake to stop it.
The three jobs above are built that way on purpose. The overdue job drafts messages and never sends one. The numbers job writes a file and prints it. The month end job writes a note for the bookkeeper and never emails the bookkeeper. That is not timidity. It is the only version of any of this that survives being put on a schedule.
Keep the rule inside the prompt as well as inside your head. The last line of the routine above says "do not send anything to anyone", and it stays there even once you trust it.
What an output you actually read looks like
If a routine runs without you watching, it has to produce something a person reads before anything leaves the building. That means three specific things, not a vague sense of a report.
A dated file on your own machine. Something like ~/Salon/reports/weekly-numbers-3-August.md, in a folder you chose, still there in March when you want to compare.
The same summary printed in the chat, so you can read the whole thing without opening the file. This sounds trivial. It is the difference between a routine you check and a folder you stop opening.
The part most reports leave out. A section that lists what could not be counted: the rows with a status nobody recognised, the completed appointments with no price recorded, the dogs with too little history to guess a cycle from, and every client deliberately skipped, with the reason. A report that only tells you the good news is not a report, it is a mood.
Then give yourself a check that takes ten seconds.
Four numbers and a total. If they do not add up, something was silently dropped, and you found out on Monday rather than at the end of the quarter.
The setting you will be tempted by, and the post that answers it
The moment you set anything to run without watching it, you will want it to stop asking you questions. There is a setting for that, and there is one option on it that has no business being on the computer your client list lives on.
We are not naming the options here on purpose. Post 19 names every one, tells you which to start on, and refuses the dangerous one out loud with Anthropic's own reasoning attached. Read it before you schedule anything, not after. This is exactly the paragraph where readers talk themselves out of it.
What this actually reclaims for you
Twenty minutes on 48 Mondays, allowing yourself four weeks off, is 16 hours a year. Right now those hours go on standing in a quiet salon deciding what to do, and then not doing it. A written routine turns all 16 into work that actually happened, and writing it took one calm Sunday.
Then the part that pays for itself. Say the overdue list turns up two dogs a week who had quietly drifted and were never going to be chased. At a $95 casual full groom that is $190 a week, and it is money your own client book already owed you. The routine did not find you new clients. It stopped you losing the ones you had, on a Monday, before the first dog was through the door.
Next in the ladder is the twelve-skill library, which is what a routine like this is made of. After that, a full week in the life shows the whole thing running Monday to Sunday. And if you would rather have reminders and payment chasing handled with no AI at all, the plain automation guide is a perfectly good place to start instead.