The other shelf
Above every grooming bench is a shelf, and it is the most honest object in the building. Clippers. The three blades that actually get used. The shears you would notice missing inside an hour. A slicker with a bent pin you keep meaning to replace.
Nothing on that shelf is decorative. Everything on it earns its place.
This is about a second shelf. There are 12 things on it, not one of them touches a dog, and you did not have to buy any of them.
The short version. A skill is a job written down: one plain text file that tells Claude Code how to do something you already do, the same way, every time you ask. These 12 pet business AI skills are the back office half of a grooming salon: your weekly numbers, overdue clients, no-shows, photos, receipts, captions, review replies, handover notes, prices, intake briefs, policies, and month end.
Every one works on a file you export or already have. None sends, posts, charges, or deletes anything. And every one carries a section on what it will never do, plus a check you can do by hand in a minute.
None of them is a program. Each is a folder holding one file called SKILL.md, written in plain English, which is why you can read one, change it, or type your own out from the shape at the end of this post. If your salon calls a full groom something else, change the word.
They were walked step by step against made up salon data first, which is where the real faults turned up. price-check was counting no-shows as services sold: on the sample export it read 23 medium full grooms where only 19 dogs had been on the table, which inflates the argument for a price rise in exactly the direction that flatters it. It now sorts the statuses the way weekly-numbers does and counts only the rows where a dog was actually groomed.
Running receipt-file a second time filed the same unreadable receipt again, so a month of repeats would have left a folder of near duplicates. That is fixed too. Two receipts printed their dates as 03/07/2026 and 04/08/2026, which a machine reading them the American way round files into the wrong month, and the rule that makes it stop and ask is the only reason they landed in July and August. A library nobody has run is a list of good intentions.
The 12 pet business AI skills, in four groups
Twelve of anything reads as a list you skim. These fall into four groups, and the groups are not equally trustworthy. Three get a longer look further down.
The money ones
weekly-numbers turns an appointments export into one page: takings, dogs seen, no-shows, busiest and quietest day, and the number that moved most. It needs 14 days of history. It will never invent a figure.
month-end gives you the month's figures, who still owes you, and a note for the bookkeeper. It needs a daily takings export and an outstanding payments export. It will never calculate GST.
price-check puts your prices against competitor prices you gathered yourself and gives the gap in dollars and against the median. It needs both lists. It will never look a price up.
The client ones
rebook-gaps finds the dogs that have quietly drifted, on each dog's own cycle, and drafts a message each. It needs a client list export carrying a last visit date. It will never state an appointment time.
no-show-follow-up lists the missed appointments, names the repeat offenders with dates, and drafts the awkward message. It needs an appointments export, plus your cancellation policy if you have one. It will never invent a fee.
The tidying ones
photo-sort matches photo times against the finish times in your appointments export and copies them into per dog, per date folders. It copies, so it never moves or deletes an original. The camera roll problem with a procedure.
receipt-file files a folder of receipt PDFs and photos by date, supplier, and total, and writes the monthly CSV your bookkeeper wants. It will never guess a total, a date, or a supplier.
The writing ones
These five end in a written judgement rather than an arithmetic result, which changes what you owe them.
caption-pack writes a week of Instagram captions from the dogs you groomed and a voice.md you write once, and stops if voice.md is missing. It will never use a dog whose owner has not said yes.
review-reply drafts replies to a batch of Google reviews, from the reviews and the same voice.md, and says which ones to leave alone. It will never name a team member.
handover-note turns a groomer's voice note transcript, and nothing else, into five sections: what was done, what to watch, what to tell the owner, what needs a vet, and what to check back. It will never diagnose, because that belongs to a vet.
intake-brief turns a two page intake form into three checks answered yes, no, or not answered, and one paragraph of 90 words at most, handling second. It will never soften a bite history, and it never reads a blank as a no.
policy-gaps marks the policies you have written against an 18 item checklist. It never drafts wording, and its own output says it is a checklist and not legal advice, which is why having them checked properly is a separate conversation.
Start with weekly numbers, because you can catch it lying
Nobody tries 12 of anything. You try one, and if it works you come back.
Start with weekly-numbers, because it is the one you can prove. Its own instructions tell you how to disbelieve it: take the busiest day in the daily table, find that day's rows in the export, add the prices on the completed ones by hand, and it must match to the cent.
The second check takes ten seconds and catches the failure that actually happens. The report prints how many rows it treated as worked, missed, cancelled, and not counted, and those four must add up to the number of rows in the file. If they do not, it has misread one of your status words, which is the difference between a Thursday that looks fine and a Thursday that lost you two slots.
That is what a good first skill looks like: something dull, with the means to check it.
Rebook gaps, and the list of people it refuses to message
The one that finds money is rebook-gaps, and the interesting part is not the drafts. It is who it removes first.
Out go anyone whose export says they opted out of messages, anyone marked do not contact, and any dog with a booking already on the calendar. Anyone who has not been in for over 12 months moves to a separate list and gets a different message, one that never mentions how long it has been, because "we noticed you have not been in for 14 months" is not a warm reopening. Any dog with fewer than three visits and no average gap column goes on a list called Not enough history, because two visits is not a pattern. A few days late is normal life; a dog is chased only when it is more than a full week past its due date.
Every removal is printed with its reason, and that list is the actual product. A message you should not have sent costs more than a booking you did not chase, which is also the reasoning behind what never to paste into an AI chat.
Review reply, and the two it will not answer at all
review-reply sorts a batch into four kinds, and the least obvious behaviour is the most valuable: two of the four get no public draft at all.
Anything mentioning a vet, an injury, a reaction, or a wound is held back, with a list of what to gather first: the appointment record for that day, who groomed the dog, any note on the booking, and any photo. The reason is in the output. A public reply can read as an admission, and you do not yet know what happened. A review that looks fake, or is simply abuse, is held back too, with what to do instead.
It checks each review against your own records where it can. In the sample data, a one star review claims a charge for an appointment that never happened, and the export shows that client with two no-shows, both unpaid. Knowing that before you reply changes how you reply, and the rules keep the fact out of the public wording, which is the harder half of the job.
Every one of them tells you what it will never do
Two sections appear in all 12 files: What it will never do, and Check this before you trust it. Neither is a disclaimer. They are the reason the rest of the file is safe to point at a real business.
Here is the thinking. A report you have no way of checking is worse than the hour it saved, because you will act on it anyway.
Adding up a fortnight of takings yourself produces a number you hold loosely and check against your memory of the week. A tidy table you did not have to add up yourself produces a number you quote to your accountant. The confidence is the danger, not the arithmetic, which is why these things get things wrong in the way they do.
So every skill prints what it found before it acts: which file it picked, which columns it mapped, which status words went into which bucket. And every one ends with a check you can do with your own eyes in a minute.
Be honest about the split, because it decides how much of your attention each one needs forever. Seven of the 12 end in something checkable: weekly-numbers, month-end, price-check, rebook-gaps, no-show-follow-up, photo-sort, and receipt-file. They count, compare, copy, or file, so the output is right or wrong and you can prove which with a calculator or by opening one folder.
The other five end in a written judgement. Their hard rules were checked, so no caption names a client, no reply names a groomer, and no handover note turns a description into a diagnosis. Whether the writing is any good is not provable, and nobody can promise you it is. Those five need your eye on the output every time, not just the first three runs.
If you edit one of these files, and you should, the never section is the part to leave alone.
The snag you will hit once, and the one you will hit every time
receipt-file reads PDFs, and on a Mac that needs a free piece of software called poppler installed. On a machine without it, the skill says so, tells you the two ways out, being photograph the receipt instead or install the tool, and sets that receipt aside rather than pretending it read something. We only know because it happened on the machine this was tested on. That is the shape of a first run: one snag, once, with a message telling you what to do.
The one you will hit every time is that your real export is messier than any sample. Your system might write a finished groom as Closed or Done, and your takings export might carry surcharges, tips, or vouchers, so month-end will report that the month does not balance on run one. That is the skill working: a month that does not balance is the finding, and the file says so rather than adjusting a figure to tidy it up.
None of these can touch your booking system
Say this out loud once, because it is the most common wrong assumption about anything with AI near it. Nothing in this library connects to your booking software. There is no live link, no login, and no token.
There is also nothing at the other end to connect to. Petboost has no setting anywhere in it that would let you plug Claude, ChatGPT, or any other AI into your account, and every skill in this library works on a file you exported yourself or a folder already sitting on your computer.
That is a limit, and it is also the safety. A skill that cannot reach your records cannot damage them. The worst outcome available to any of the 12 is a wrong figure in a report you were told to check, or a draft message you decide not to send.
In most systems the exports sit a few clicks away, under Reports. In Petboost they come from the Export Your Data screen under Billing, which is a CSV per data type rather than one button. Get the appointments export first, because six of the 12 read it.
Writing your own pet business AI skills from here
Nothing stops you typing one out yourself in a text editor. All 12 follow the same shape:
The description decides whether it fires when you ask in plain words, so write it with the words you would use at 6am, not the words a manual would use. What it does is numbered steps in plain English, and What it will never do is a line you would say to a new starter on their first morning.
Each skill is a folder with SKILL.md inside it, dropped into the skills folder Claude Code already reads, and the guide to building one from nothing covers where that goes. Leave Claude Code in the setting it arrives in, where it asks before changing a file. Nothing here needs more than that, and what to allow and what never to is its own conversation.
What this actually reclaims for you
The admin evening in our worked example is two hours, twice a week, which is four hours Steph is trying to get back. She is a composite we follow through this series, and her numbers are the sort you would recognise rather than a real client's books.
Say half of those four hours is the countable half: the weekly figures, the receipts, the overdue list, and the month end. That is two hours a week, every week, on jobs you never explain again, because the explanation is sitting in a file. And the first Groodle that comes back off the overdue list rather than drifting to the salon two suburbs over is $95 you had written off.
That is the honest ceiling. Not a business that runs itself. Two hours, and a shelf of tools you can read.
Once one of them works, the question stops being what they should do and becomes when they should run, which is the Monday morning routine. And to see the whole lot in motion across a real week rather than described one at a time, there is a week in the life of an AI back office.