Skills · digital agency
The same method applied to a business with almost nothing operationally in common with a donut shop: departmental P&L, revenue recognition off a billing sheet, monthly billing and collections, and a partner deck every month.
Most of these now post straight to QuickBooks 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. The one agent that writes to QuickBooks on its own can only create bills from invoices a teammate has already approved — it can’t edit, pay or delete anything.
Refreshes the monthly financials workbook from QuickBooks, then builds the ~31-slide partner deck out of it — every dependent tab updated, headcount reconciled against payroll, and four partner personas pressure-testing the deck before it reaches a real partner's inbox.
Builds a monthly cash basis P&L with client ad spend taken out of both sides — the client money coming in and the media bills going out. It rebuilds every bank transaction from the ledger, one row per transaction per account, and each account has to tie to its own balance sheet change to the penny before a single row gets classified.
Written rules settle the judgment calls, because judgment that lives in someone's head gets applied differently in March than in August. Interest counts when the bank deposits it, so moving it later never counts it twice. Commission counts when it moves into operating, since until then it is sitting mixed in with client money.
Builds the semi-monthly payroll journal entry as a QuickBooks import, pulling payroll straight from the provider's API rather than from a manual check entry. It reads paystubs, fees and member departments, converts cents to dollars, and aggregates earnings, deductions, employer contributions and workers' comp into a balanced entry that credits the operating bank account for the invoice total. Non-executive rows map to salary and wages by department; executive rows map to their own accounts and classes.
It exists because the native payroll-to-QuickBooks sync maps executive compensation wrong. The import format is a spec arrived at after six failed attempts — leaf account number plus full qualified path, currency on every row, fully qualified classes, and a plain numeric journal number repeated on every line.
Turns the month's billing sheet into a balanced draft revenue-recognition entry, then converts that draft into an import file once it's been reviewed. It reads only the rows tagged for rev rec, finds columns by header name rather than position because they move between months, and reads the human-written finance notes to derive hours and rate per performing department. It debits where the revenue was billed and credits the department that did the work, carrying the client tag, a derived class, and the originating row numbers.
Billing week is where the money is either captured or quietly lost. These four cover the month's biggest billing run, the invoice before it goes out, the chase when it comes back late, and the cash when it finally arrives.
Runs the month-end paid media management fee billing end to end. It reconciles every client's ad spend against the spend source to the penny, independently recomputes every fee in code from the pricing rules — tiers, minimums, caps, and the aggregates where several brands bill as one family — and reports any row where its own calculation disagrees with the sheet.
Then it drafts each invoice by cloning that client's prior-month one, so terms, billing contacts, purchase order fields and line formatting carry forward instead of being retyped. It shows me a table and creates nothing until I approve it. After approval it creates the invoices, never sends them, writes each invoice back to the sheet as a link, and checks the posted result against what was asked for — date, amount, and description, line by line. The last step ties the fee income account to the sheet total within a penny.
The guardrail that matters most is a blank cell. Where a family of brands bills on aggregate spend, the parent row has to stay empty; a number there gets counted twice, which inflates total spend and — because the tiers run the other way — quietly drops the whole family into a cheaper rate. Nothing about the sheet would look wrong, so the skill checks for it every month.
Reviews every invoice created during monthly billing against the billing sheet and the ledger, using a fifty-three point checklist built up from real billing runs and the bookkeeper's own rules. It reads the invoice lines, the project manager's notes, the linked estimate and last month's row, then redoes the hours math and shows the arithmetic inside each finding. Findings land on a review page in three tiers: block, fix before close, and advisory.
Before it calls anything a fix, it does one extra read that could disprove the finding. A reviewer who cries wolf gets ignored, which is worse than no reviewer at all.
Drafts a past-due chase for one client: a short email and a one-page statement of every open invoice. The email carries two numbers — total outstanding and amount overdue — and nothing else. It never says when an invoice passed terms, because invoices sometimes go out late and that framing starts an argument you cannot win. The statement ages each invoice against the due date printed on it, so it matches the document the client is actually holding.
Records a batch of customer checks deposited by phone. It reads every check — payer, check number, the amount in numbers and again in words, memo, and any remittance stub — confirms the two amounts agree, and matches each check to open invoices. A check with no invoice number only gets applied when the open invoices on a single project add up to it exactly. Anything else gets flagged rather than guessed.
Then it stops and shows the plan. After I approve, it re-checks the invoice balances in case someone recorded a payment in the meantime, posts one payment per check, and reads each one back. It also crops each check out of the scan and names the image by invoice, payer, check number and amount.
Ties one month's column of the prepaid expenses amortization schedule to the general ledger and explains any variance transaction by transaction. It pulls month-end balances from the accrual balance sheet, compares the schedule against the ledger, and walks back through prior months to find where the break started — then queries journal entries, bills and purchases to isolate the lines that actually drove it.
The usual culprit is timing on new subscriptions: a renewal bill or card charge dated late in the month hits the ledger, but the schedule keeper parks the setup in the following month's column. A few cents of residual is normal penny rounding on the monthly amortization entries and isn't worth chasing.
Builds the monthly balance sheet reconciliation. It copies last month's workbook, loads the balance sheet in its export format, and reconciles each supporting tab: receivables and payables aging, undeposited funds, fixed assets, accrued expenses, and unearned revenue by customer and project. It ties the linked schedules for prepaids, PTO and the ad spend liability without editing them.
It checks completeness the hard way: every account carrying a balance has to have a row on the summary, because a lookup that defaults to zero will quietly hide one that doesn't. Unearned revenue gets a year-to-date check row by row, since an entry on the wrong project still leaves the total tying. The last step compares every account to the trial balance in code.
Rolls the client ad spend liability reconciliation into a new month inside the main workbook. It duplicates last month's tab, pulls the month-end balance sheet into a companion tab, and rebuilds the invoice lists, the vendor bills coded to the liability accounts, and each client's balance from their own spend tracker. Before writing anything it simulates the tracker lookups in code, so it knows what every cell should say before it says it.
The tab explains its own cash variance. It lists tracker spend not yet billed and client payments still sitting in operating, and keeps going until the line labelled unexplained reads zero.
Two skills, deliberately split. The expense one drills to vendor; the revenue one drills to client. If the account is income, the revenue skill supersedes.
Runs a month-over-month flux on any expense account or account group, drilled to vendor and transaction level. It maps a plain-English category to real accounts, pulls the P&L for both periods to set the reconciliation target, then queries all three sources that post to expense accounts — card charges, AP bills, and journal entries carrying the monthly amortization of annual prepaid software. It handles the credit flag on purchases, signs journal entries by posting type, and infers vendors for journal entries from line descriptions.
The hard rule is that nothing is produced until all three sources reconcile exactly to the P&L. Amortization entries are the source most easily forgotten, which is why the reconciliation is a hard gate rather than a check.
The income counterpart, drilled to client rather than vendor. It maps plain-English revenue terms to account families, sources by-customer revenue from a transaction-detail export, pulls client names out of commission journal-entry descriptions where the name field is blank, and rolls sub-jobs up to the parent client against a canonical list. It can corroborate a movement with logged hours rather than leaving it as a guess.
Each period has to tie to the penny before the workbook gets built. A zeroed-out client isn't automatically a loss and a new one isn't automatically a win — the hours tab is there so you can tell which.
The other side of the same method — seventeen skills for a nine-location food business →