How to run your firm’s month-end close in Claude, not a generic checklist.
Claude can run the close that already lives in your firm. Your tiers, your materiality, your capitalization, your trigger phrase, the package that has to land in the client folder. A fractional CFO Claude QuickBooks close is the same sequence with a different artifact at the end: stoplight, board pack, WIP by project, not a vanilla P&L. It cannot invent any of that, and it will try if you let it.
Split the book before you write a prompt.
A generic close checklist is how you get a confident, wrong month. I have watched it happen on our own book at AG Accounting: the model finishes, the P&L looks tidy, and a restatement is sitting in last month’s issued pack that nobody compared against. An outsourced CFO month-end close fails the same way, then the 13-week cash forecast inherits it.
Claude can run the close that already lives in your firm. Your tiers, your materiality, your capitalization, your trigger phrase, the package that has to land in the client folder. A fractional CFO Claude QuickBooks close is the same sequence with a different artifact at the end: stoplight, board pack, WIP by project, not a vanilla P&L. It cannot invent any of that, and it will try if you let it.
Run one Claude project per client, against live QuickBooks Online or Xero. Procedure in the skill file. Client facts in the client file. A trigger phrase the whole team says the same way. Master prompts inside the 8,000-character firm-section ceiling, or split them. Save last month’s balance sheet and P&L before you start, or you cannot see a restatement. Flag exceptions first, then drop the preset package in the folder and mark the file done on a list the partner can actually read. Prompt quality is the close-time variable. Stop if the skill is stale. Do not dump unrelated clients into one project.
The control design (SOP library, precedence, versioning, named acknowledgement, the rule that publishing does not remotely update anyone’s installed skill) lives on firm-controlled AI instructions. Read that when you are building the library. Come back here when it is the last week of the month.
This is not AI bookkeeping. Nobody is hiring a categorizer. Statements in, AJEs proposed, exceptions flagged, package in the folder, a person on every write. Same headcount. The hours you get back go into the sentence the client will pay for.
Around client twelve, firms discover that cash-basis startups and full accrual shops with AR, AP, and inventory cannot share a brain. We learned that the expensive way. One heroic close prompt does not survive contact with the actual book. A fractional CFO book of eight to forty files hits the same wall faster, because the files are less alike: a construction WIP next to a SaaS accrual next to four entities under one founder.
Tier the way you already staff it. Ours looks roughly like this. Use yours.
- Tier 1-2. Cash or modified cash. Bank and card activity, a short vendor list, few or no accruals. Close is coding, recon, a short exception scan, and a package.
- Tier 3. Accrual with recurring AJEs (prepaid, deferred revenue, straight-line items). Recon still matters. Judgment entries show up every month.
- Tier 4-5. Full AR/AP, inventory or job costing, multi-entity, or a board pack. Materiality is tighter. Capitalization is not a default. The package is not a vanilla P&L.
Three master prompts usually beat five. Five tiers of clients can still share three close brains if the client file holds the facts: basis, thresholds, who gets 1099 treatment, which accounts are never auto-cleared. Procedure in the skill. Facts in the client file. Mix them and a capitalization change on one client poisons the prompt for everyone else.
The raw asks we send are in the prompt library. Steal the structure. Leave our materiality alone.
Materiality and capitalization belong in the client file.
Claude will invent a threshold if you let it. Not because it is sneaky. Because the prompt said “use professional judgment” and the file was silent.
Write the numbers.
- Materiality by tier, in dollars, on the client file. A $500 miss you would ignore on a cash-basis shop is not ignorable on a board-pack client.
- Capitalization by tier, in dollars. $2,500 is common. If you use $1,000 on one book and $5,000 on another, that is a client-file fact, not a debate for the model. Missing number: stop and ask. Do not guess.
- Dormant-account rules, 1099 flags, class requirements, and “never code payroll from the bank line” belong here too.
A stalled close is cheaper than a confident close that capitalized a laptop on the P&L because last quarter’s chat “remembered” $2,500. I would rather a staff person ping me at 6pm than find that in the issued pack.
The trigger phrase is how the team starts the same close.
Two words. 32 characters. It invokes your close rather than the shipped one. Firms treat this like branding. It is not branding. It is how you stop three people from running three different checklists on the same Tuesday. A solo fractional CFO still needs it, because the contractor covering board week will otherwise type “do the close thing” and hit last month’s default.
Until the custom phrase is published into the skill everyone has installed, pick a standing phrase and write it on the who’s-done board. “Close the month” is enough. Do not leave people typing “run the wizard” or “do the close thing.” Those hit whatever default the last person left in the project.
Publishing a new firm section changes the next download. It does not reach into Claude and update anyone. Treat a publish like a release. Control design is on firm instructions.
8,000 characters is a ceiling, not a dare.
The firm instruction section is capped at 8,000 characters. Stuff a 14,000-character master prompt into the client file, or paste it into every project, and the model skips the middle, ignores the recon rule you buried on page four, and still sounds sure. Close week does not need more confidence.
A twenty-page close document does not belong in the skill as twenty pages. Split it:
- SOP (human). What a person does. No tool identifiers. Lives in SharePoint or the handbook. Survives a vendor renaming a button.
- Firm section (AI). What the model does, inside 8,000 characters, with precedence over the shipped recipes. Trigger phrase, order of operations, what to stop for, what never to invent.
- Client file (facts). Basis, thresholds, quirks, last month’s package location. No procedure.
When a firm sends us a 14k block and asks whether it can hold that, the answer is trim and split. We can help you cut. The model cannot. A long prompt that almost fits is worse than a short one the team actually runs.
Save last month’s balance sheet and P&L or you are flying blind.
Most firms skip this because pulling a trial balance “as of today” feels like the same report. It is not.
A trailing TB pulled today cannot tell you last month’s issued revenue moved. Prior-period restatements, reversed accruals, a helpful staffer recoding a class on the 12th after the PDF went out: all of that is invisible if the only numbers you hold are live. The live TB absorbs it. Your issued PDF does not. A fractional CFO who then builds a 13-week cash view off the live TB has just modeled the restatement.
Before the trigger phrase:
- 1. Confirm last month’s issued balance sheet and P&L are saved where the skill can see them. Saving them in Numbers Game (the explicit saved-report exception) or dropping the PDF into the client folder both work. Re-pulling is not saving.
- 2. Tell Claude to compare this month to that package, not to a fresh trailing twelve.
- 3. If last month is missing, stop. Do not close against a reconstructed memory of what you sent.
Client reporting is where storage and retention are stated precisely. A saved monthly pack and a one-off analysis should not share a retention setting just because both are “reports.”
Exception flags first, then the stoplight package.
A tidy P&L is not a close. A close is a person knowing which files need a conversation. Fractional CFO client reporting starts here: the color, then the letter, then the 13-week.
Ask for flags before the package. Four that pay for themselves on a multi-client book:
- A balance-sheet account that is negative and should not be (cash, inventory, a prepaid that flipped, retained earnings underwater for a reason that is not the close).
- Losses for more than three months.
- A sudden increase in debt.
- Owner draw high relative to that client’s pattern, or relative to the wage the owner is actually taking.
No flags: send it to the preset package. One flag: the package still runs, but the board does not say “good to go.” It is a review item with a sentence attached.
Then the boring part you already standardized:
- Run the financial package for that tier (P&L, balance sheet, cash flow, whatever they get). Construction clients get WIP by project. Board clients get the stoplight. Outsourced CFO month-end close often adds budget-to-actual on the same pass.
- Save it in the client folder, the same place the team already looks. SharePoint, TaxDome, Drive. The folder is yours. The connector is not your DMS.
- Mark the client done on one list the partner can read without opening forty Claude projects.
Traffic lights beat Slack noise. Most CPA firms on this product live in Teams and TaxDome. Fractional CFOs live in Drive and email. Do not wait for a Slack dashboard to make the close real. A SharePoint list, a TaxDome job status, or a sheet with client, tier, flags, package saved, reviewer is enough. Slack is optional. The board is not.
Prompt quality is the close-time variable.
Same product, same connector: one file closes in about fifteen minutes, another takes thirty or forty because the prompt was vague. The model was not worse on the slow file. The instructions were.
Weak close prompts look like this:
- “Review the books” instead of naming the recon, the uncategorized queue, the AJEs, and the flags.
- Claude Googles a new vendor and invents treatment. Research is useful. Invented treatment is not. Unknown vendor: stop and ask, or follow the firm’s new-vendor SOP.
- A request for a “clean” package, which is how you get cheerleading and a missing accrual.
- A different paste in every project, so you cannot tell whether the slow file is the client or the drift.
Refine on one representative client per tier, then ship that version. Do not customize in the chat and forget to write it back. The unglamorous hour you spend on the prompt is the hour you stop spending on every subsequent close.
Stop if the skill is stale.
Claude will happily run last month’s close against this month’s books using v3 of a prompt you replaced with v6. It will not volunteer that it is stale. People will not check. Close week is loud. Version numbers are quiet.
First operation: version check. If the installed skill or the MD file in the project is not the current published version, stop. Download again. Choose Replace. Refresh the connector and the skill together. A few minutes. A release, not a suggestion.
You should be able to see which projects are behind without opening them. Until that view exists, the who’s-done board needs a skill-version column, or you are certifying closes you cannot reproduce.
This is not the migration week. Migration is connecting files. This is keeping the procedure current after they are connected.
Do not batch unrelated clients into one Claude project.
All the Tier 1 cash-basis shops look the same. One project, one prompt, a loop. We have been asked. The answer is no.
Grouping here means parent and child, related entities, combined reporting. It does not mean every cheap client in a pile. Claude will book into the wrong company if the project can see more than one unrelated ledger. That mistake is expensive and quiet. One project per client. The multi-client tools guideis the operating-stack version of the same rule.
Related entities are a different case. Multi-entity QBO (four entities, the tab-switching that made the fCFO hire you) is a real workflow. Combined reporting across connected companies is allowed. One project per entity still. Read multi-entity reporting before you promise a consolidation. Combined is not consolidated.
Scheduled work, if you want it, is Claude scheduled tasks today. ChatGPT can run the same connector. It does not currently give you the same scheduling story. Do not design month-end around a scheduler you do not have.
What still needs a person.
Human approval before writes is not a loading screen. A person still:
- Approves every proposed journal entry, categorization, and recon adjustment.
- Decides whether a flag is a client conversation or a pass.
- Signs the package. Claude drafts. Your name goes on it.
- Refuses to close if last month’s pack is missing, the skill is stale, or a threshold was never written down.
- For a fractional CFO: refuses to model until the audit is done. The 13-week is not a substitute for the close.
The connector will not permanently delete ledger records, change platform settings, initiate payments, touch payroll, mark QBO reconciled (it can produce a rec report), read the QBO bank-feed API, or ingest a Word close template with row shading. Work around those on purpose. Do not write them into the SOP as if they work. Product-level exclusions live on features.
A close sequence you can steal tomorrow.
Use this as the runbook. Put the firm-specific numbers in the client file.
- 1. Confirm skill version. Stop if stale.
- 2. Confirm last month’s issued BS and P&L are saved. Stop if missing.
- 3. Confirm accounting basis, materiality, and capitalization are on the client file. Stop if missing.
- 4. Say the trigger phrase.
- 5. Uncategorized and bank/card activity, proposed, not posted.
- 6. Recon against the statement you actually have. Do not plug. Do not pretend a missing statement is a rec.
- 7. Recurring AJEs for that tier, proposed as a set.
- 8. Exception flags (negative BS, three months of losses, debt spike, owner draw).
- 9. Package into the client folder (stoplight, WIP, budget-to-actual, whatever that tier actually gets).
- 10. Person reviews flags and proposed writes, approves, marks the board.
- 11. Only then: 13-week cash, board commentary, or the model. Not before.
That is a close. It is not a chatbot session that happened to mention QuickBooks.
Close questions, answered.
Book a walkthrough on a real close.
Bring one client file, last month’s package, and the close prompt you actually use. We will run the sequence, not a demo script.