Skills · Union Square Donuts
A skill is a written procedure Claude follows — a workflow described once, in enough detail that it runs the same way every time. These seventeen cover most of the recurring finance work at a nine-location food business: two point-of-sale systems, cash counted by hand at every store, tips owed across both, bank and balance sheet reconciliations, the recurring journal entries, and a meals tax filing. Together they're what made a twenty-hour month-end close into a three-hour one.
These used to end in an import file I ran myself. They post straight to the books now, through an MCP connection — but only after I approve that specific entry. The skill shows me the draft, I say yes, it posts, and then it reads the entry back to prove it landed the way it was built.
Builds a balanced, QuickBooks-ready journal entry for multi-location Square sales over any date range. It pulls live from the Square reporting API, excludes third-party delivery apps, rolls locations up to accounting classes, maps product categories to revenue accounts, splits cash from card settlement, and books service charges and over/short — then writes a balanced entry and posts it once I've approved it.
This one took the longest to get right, and not because of the code. Five commercial products were demoed first. None of them got the mapping exactly right, and none were as fast.
The food hall runs on a different point-of-sale system, so it reports separately. This pulls a sales-and-payment summary day by day — orders, gross, discounts, net, tax, tips, total collected — with a payment-by-tender breakdown underneath and totals that tie out.
Takes the payroll reports package as it arrives — a PDF — plus the account mapping, and produces the journal entry with payroll detail broken out by department. At month end it also produces the accrual.
Tips arrive from two different point-of-sale systems across every location. This pulls both, for whatever dates you specify, and reconciles them against each other so the two systems agree before anyone gets paid.
Two separate reconciliations that both live in the gap between a month ending and the money arriving: physical cash collected but not yet deposited, and card payouts on the month's revenue that settle to the bank the following month. They're kept as distinct reconciliations in one workbook rather than netted together, because they break for different reasons.
Maintains the fixed asset register and the depreciation schedule behind it — additions, disposals, the monthly depreciation run, and the tie-out back to the fixed asset accounts in the general ledger.
The monthly tie-outs. Each of these ends in a number that has to agree with something outside the ledger — a bank statement, a drawer count, a signed lease — and says so plainly when it doesn't.
Reconciles the month across three operating accounts at two banks, reading the statements as they arrive as PDFs rather than waiting on a feed. It ties each account to the ledger and delivers all three as one formatted workbook, filed to the project and to Drive so the backup is where the next person would look for it.
Reconciles the cash that never leaves the building — safe and drawer balances at every location — against the ledger. It reads the counts by location from the cash tracker, compares them to what the books say should be sitting there, and proposes the entry for the difference.
Nine sites each holding float is nine ways for a balance to drift quietly, and a drawer that has been wrong since March is a much worse problem than one that was wrong last week.
Rolls the security deposit schedule forward to month end, by location, and ties it to the ledger. Deposits sit on the balance sheet for years and get touched twice a decade, which is exactly the kind of balance that turns out to be wrong during diligence.
Produces the fully-loaded P&L by location. A store's own P&L is the easy part; the useful part is pushing commissary and overhead costs out to the stores that actually drive them, following the management allocation methodology rather than splitting evenly.
Once a month closes, its actuals replace the forecast in the three-statement model and the projection rolls forward. Doing this by hand is where models quietly drift out of agreement with the books.
Four entries that recur every month and were each, at one point, something I rebuilt by hand because the rule behind it lived in my head rather than anywhere a system could read.
Handles the month-end catering cutoff. Catering gets invoiced and paid on its own schedule, which has nothing to do with when the food was actually delivered, so it reads the invoices by service date and sorts them: unpaid invoices for work already delivered accrue to receivables, invoices already paid for work not yet delivered get deferred to customer deposits.
Both entries reverse the following month, so the cutoff corrects itself rather than needing to be remembered and unwound.
Online catering orders land in the ledger as income with no location and a generic sales account, because the ordering platform has no concept of either. This moves that revenue to the right location and the right sales account every month, so location reporting — which is how the business is actually managed — isn't quietly missing a channel.
The point-of-sale system collects a kitchen fee and lands it in the same bucket as sales tax, where it sits looking like money owed to the state. It isn't — it's income. This moves it out of the tax liability and into income every month, which keeps the amount actually remitted to the state equal to the amount actually collected for it.
Accrues monthly interest on outstanding loans, day-weighted so a loan funded mid-month accrues from the day the money arrived rather than from the first of the month. Loan interest is the sort of accrual that gets skipped for three months and then booked as a lump, which makes every monthly P&L in between wrong.
Reconstructs what was actually on the shelf at any store at any past moment, by taking current inventory and replaying the change log backwards to the time in question. Answers the question you can't otherwise answer after the fact: how many did we have left, and of what.
Vendor credit applications all ask for the same thirty fields in a different order — entity details, banking, trade references. This fills them from a saved business profile instead of retyping them each time a new supplier wants net terms.
The other side of the same method — thirteen skills for a digital agency →