Nothing to hand? Load the — a close that is genuinely fine — the , where one unconfirmed counterparty balance holds ten tasks behind it, or the , where the target is arithmetically reachable and the execution is eight days behind it. All three replay a saved run for free. Or press to watch the checker name its findings with no model call at all.
A close calendar is business days, and business days are not arithmetic on a date
T+3 is the third business day after the period ends, and it moves when a company holiday lands in the middle of it. So the engine builds the calendar from your period end, skipping weekends and every holiday you declare, and every deadline below is measured against the day it actually falls on. Overdue is counted in business days too, because "two days late" over a weekend means something different from "two days late" mid-week.
The dependency ladder is what makes a checklist mean something
Every task lands on one of five levels — post the source entries, reconcile the subledgers, reconcile the balance sheet, build the statements, review and lock — either from a Level column you supply or by matching the activity against the canonical checklist, and the engine tells you which of the two it used for every row. From that structure it can answer the question a tick-box list cannot: this one open task is holding ten others, and that one is holding nothing.
The quiet failure is the task that finished too early
A trial balance signed off while the bank reconciliation was still open was computed on numbers that have since moved, and nothing in a spreadsheet will tell you. The engine finds every completed task sitting above open lower-level work and names exactly which tasks it jumped, so the rework is scoped now rather than discovered at management review. The same pass checks your list against the canonical activities and separates the ones that are genuinely required from the ones that simply do not apply to your entity.
The metered pass is judgement, and it is held to the measurement
What a model is for: what a blocker actually costs in days, what to move and where, what to raise at the fifteen-minute standup, what to escalate and to whom, and what to automate before next month. It returns exactly one review per flagged task, and the app counts them — a task reviewed twice fails as loudly as one left out. Each severity is checked against the measured downstream count and days late, every date it cites must exist in the calendar, every person it names must own a task on your list, and its posture is set beside the rule's rather than merged into it.
What does Close Desk measure that a checklist in a spreadsheet does not?
The dependency structure underneath the checklist. A spreadsheet knows a task is unticked; it does not know that the one unticked task at level two is holding ten tasks at levels three through five, that the trial balance was signed off while the bank reconciliation was still open, that the target close is shorter than the dependency chain can physically run, or that four of the tasks due on the same day all belong to the same person.
Do I have to sign in?
No. The parsing, the calendar, the dependency ladder, the downstream counts, the overdue arithmetic, the out-of-sequence detection, the checklist coverage, the materiality judgement, the owner load, the posture rule, the whole status pack, the task CSV, the calendar file and the measurement JSON all run in this tab with no account and nothing charged. Only the judgement pass is metered, and the three bundled closes replay a saved run for free.
Why does it flag tasks that are already marked complete?
Because a completion that ran ahead of its own inputs is the quietest way a close goes wrong. The dependency map says level three cannot legitimately finish while level two is open. The engine names each one and lists which open tasks it jumped, so you know what has to be redone rather than finding out at management review.
What is checked about the AI pass?
Every flagged task must get exactly one review — a task reviewed twice fails as loudly as one left out. Each severity is compared against the measured downstream count and days late, so calling a task that blocks ten others low severity fails. Every close day and date cited must exist in the computed calendar, every person named must own a task on your list, a task measured as complete cannot be reviewed as an open problem, and the pass's posture is set beside the rule's rather than merged into it.
What does a run cost, and what if my balance is low?
A worst-case amount is reserved and only what the run uses is charged. Before you submit, the estimate is checked against your balance and the answer is shown at the run button rather than after it: signed out, it says to sign in and why; under the model's minimum, it names the exact shortfall and offers a top-up; and a reply cut short by a low balance says so rather than presenting itself as a complete pack.
Where does the pack go afterwards?
Out of the browser, with its audit attached. Markdown for the status mail, a CSV of every task, missing activity, reconciling item, check and verification line for the spreadsheet, an .ics of the close days for the team calendar, and the whole measurement as JSON. Closes are saved to your account rather than to one browser, which is what makes the comparison this app exists for possible: this month against last month, and whether the same thing slipped again.
A derived work of @anthropics/close-management: its dependency map, its five-day and accelerated close calendars, its pre-delivery checklist, its status definitions and its bottleneck catalogue are what this app implements and measures. The skill asks a model to run the close; this computes the structure and holds the model to it. Close Desk is a close-management aid, not financial, accounting, audit or tax advice.