Money & Numbers
Ask clients for missing files, explain what changed in the numbers, and keep a clean review trail — without trusting made-up figures
Anyone can use this page. Copy the prompt and try it today.
$9.99
Coming soon
Everyday AI
Client requests, change notes, and the review trail.
You stay the editor. The model drafts. You decide.
$9.99 · 20 pages
Money & Numbers · accountants · never invent numbers
CONTENTS
Read pages 3 through 6. Then copy one cream box.
WHAT THIS BOOK IS
Thursday at 5:48, Priya has three close notes open. One says only “timing.” Another has a number with no source. The third is an email she has not sent.
This book helps with one small accounting job at a time. Pick a page. Copy its cream box into ChatGPT, Claude, or Gemini. Replace the brackets with redacted facts you are allowed to use.
The model can arrange your words. It can turn a list into a table. It can point out a missing fact. It cannot see your ledger (the official list of accounts and balances), prove a balance, approve an entry, or know the right accounting treatment.
You stay the accountant. The model drafts. You trace, recalculate, correct, and approve through your normal process.
Hard nos: No invented balances, dates, causes, citations, tax amounts, approvals, or procedures. No confidential records in an unapproved tool.
Then do this: Choose one low-risk draft this week. Start with redacted notes, not a live client file.
WORDS TO KNOW
Wednesday at 7:10, Noah opens a prompt and sees five labels. They are simple once each has a job.
A ledger is the official list of account activity. A reconciliation means matching two lists of money and explaining differences. A source label such as GL-1 points back to a record.
Hard nos: Do not treat labels as proof. A well-shaped answer can still be wrong.
Then do this: Add source labels to the notes for your next draft.
WHAT’S INSIDE
Monday at 4:35, Elena wants help with everything at once. She gets a longer answer and a larger review problem.
We will not use a chatbot as the ledger, calculator of record, standards source, tax adviser, control owner, evidence file, or final reviewer.
Before pasting, check firm policy and the tool’s data rules. If safe redaction would ruin the task, do not paste it.
Hard nos: No live ledger dumps, bank files, payroll details, tax IDs, passwords, or restricted audit evidence.
Then do this: Name the job in one sentence. If there are two jobs, split them.
WHAT THE LABS FOUND
Friday at 3:20, Marcus asks whether research proves an AI-written memo is right. It does not.
OpenAI: Writing is about 40% of work-related ChatGPT messages. About two-thirds of those writing messages ask the model to change supplied text. This supports an edit-first habit.
Anthropic: Putting long reference material before the question improved some complex, multi-document test results by up to 30%. This is not an accounting assurance result.
Anthropic’s AI Fluency Index: People who iterate also show more clarification and checking habits. The link is correlational. It does not prove that more prompts cause better work.
Google: Useful prompt parts include objective, persona, context, constraints, and response format. Our version is Role, Task, Details, Limits, and Format.
Hard nos: None of these findings proves that an AI draft is complete, compliant, or correct.
Then do this: Use research to improve the request. Use records and professional review to test the result.
START SMALL
Tuesday at 6:12, Elena asks for a variance note. The answer blames seasonality. She never supplied monthly history.
COPY THIS
Role: You are a drafting assistant. I remain the accountant and reviewer.
Task: Structure a draft using only my supplied records.
Details: Purpose: [close note or client email]. Period: [period]. Approved facts: [redacted facts]. Source labels: [TB-1, REC-2]. Unknowns: [list].
Format: 1) evidence table; 2) short draft; 3) open questions.
Hard nos: Do not invent numbers, dates, causes, citations, tax treatment, or completed procedures. Keep facts, my calculations, assumptions, and open points separate.
The blind-paste risk is a smooth answer that fills gaps with likely-sounding details.
Then do this: Trace every sentence to a source label. Delete or mark any line that cannot be traced.
CLIENT REQUEST
Thursday at 10:05, Marcus needs three files from café owner Ana. His first email is full of accounting terms. Ana sends the wrong month.
COPY THIS
Role: You are a clear client-email editor.
Task: Turn my approved missing-items list into one courteous request.
Details: Client label: [Client A]. Period: [June 2026]. Needed items: [June bank statement; June loan statement; June merchant report]. Approved reason: [reason]. Deadline: [date]. Secure channel: [portal].
Format: Subject; short opening; numbered list with item, period, and file type; deadline; secure-upload reminder.
Hard nos: Do not add requests, imply wrongdoing, promise completion, or include private data.
The blind-paste risk is scope creep. The model may request extra records because they are common elsewhere.
Then do this: Compare the email with your approved request list. Check every item and period.
VARIANCE NOTE
Monday at 7:02, Talia sees software expense jump. The jump is bigger than her notes explain. One invoice (a bill for goods or services) supports part of the change. Her notes do not explain the rest.
COPY THIS
Role: Help draft a variance note. A variance is a change between two amounts.
Task: Describe the supplied movement. Keep proven drivers separate from open questions.
Details: Current: [$___, TB-C]. Comparison: [$___, TB-P]. My checked difference: [$___ and ___%]. Supported drivers: [INV-7, $___]. Unexplained remainder: [$___].
Format: Inputs; evidence table; explanation under 80 words; follow-up questions.
Hard nos: Do not calculate missing figures, call a link a cause, use “timing” without evidence, or invent events.
The blind-paste risk is a tidy story whose parts do not add to the total change.
Then do this: Recalculate outside the chat. Confirm that the supported drivers and open remainder match the total movement.
CLOSE CHECKLIST
Wednesday at 4:50, Jorge inherits an approved close process written in eleven paragraphs. He needs a checklist, not a new control system.
COPY THIS
Role: You are a careful procedure editor.
Task: Convert the supplied approved process into a checklist without changing it.
Details: Approved process: [redacted text]. Roles: [preparer and reviewer]. Due points: [dates]. Required proof: [schedule, attachment, sign-off].
Format: Table with order, task, owner, input, proof, due point, reviewer, and status. List anything not stated.
Hard nos: Do not add controls, remove steps, assign unnamed people, invent dates, or make steps optional.
The blind-paste risk is silent process redesign.
Then do this: Read the source and checklist side by side. Ask the process owner about any unclear order or owner.
RECONCILIATION NOTE
Friday at 6:25, Sasha’s reconciliation works, but the note says only “difference investigated.” Next month, no one will know what she checked.
COPY THIS
Role: Draft reconciliation notes from supplied evidence.
Task: Record what was compared, the differences, the steps actually performed, and open items.
Details: Account: [label]. Period: [date]. Ledger balance: [$___, GL-1]. Supporting balance: [$___, SUP-1]. Difference items: [item, amount, source]. Steps performed: [steps]. Open items: [owner and date].
Format: Purpose; sources; steps; results table; conclusion placeholder; reviewer questions.
Hard nos: Do not say “reconciled,” “complete,” or “no exceptions” without support. Do not invent a step.
The blind-paste risk is claiming that normal procedures were performed when the file does not show them.
Then do this: Match every procedure sentence to saved proof. Keep the conclusion blank until checks are complete.
JOURNAL ENTRY
Tuesday at 5:40, Lena has an approved entry with the note “fix accrual.” Six months from now, that line will not explain the event.
COPY THIS
Role: You are a journal-entry note editor, not an approver.
Task: Draft a clear narrative from approved support.
Details: Entry ID: [JE-104]. Approved debit and credit summary: [accounts and amounts]. Event: [verified fact]. Period: [period]. Support: [labels]. Approval status: [status]. Reversal: [verified instruction or not supplied].
Format: Narrative under 70 words; support list; open questions.
Hard nos: Do not choose accounts, change amounts, claim approval, invent purpose, decide tax treatment, or create reversal instructions.
The blind-paste risk is a fictional business reason built around a sparse note.
Then do this: Match account names and amounts to the approved entry screen. Use the normal approval route.
REVIEW NOTE
Thursday at 3:15, a reviewer asks Dev, “Why is this still open?” Dev reads blame into the sentence and writes a defensive reply.
COPY THIS
Role: Help me answer a reviewer’s note in neutral language.
Task: Separate the question, supported answer, missing proof, and next action.
Details: Review note: [redacted note]. Supported facts: [facts and labels]. Missing proof: [list]. Proposed action: [action]. Target date: [approved date].
Format: Restated question; direct answer; source labels; open point; next action. Maximum 120 words.
Hard nos: Do not claim completion, blame anyone, judge the reviewer, invent a date, or close the note.
The blind-paste risk is turning an open item into a reassuring but false “resolved” statement.
Then do this: Ask whether the reply gives proof or only calm wording. Keep the note open while evidence is open.
GUIDANCE SUMMARY
Monday at 8:05, Mei has a long standards document. She needs a starting map, not a citation produced from memory.
COPY THIS
Role: Summarize only the guidance text I supply for professional review.
Task: Extract relevant requirements and preserve locations I can verify.
Details: Question: [issue]. Framework: [specified]. Effective date: [known]. Source text with page or paragraph labels: [text]. Redacted facts: [facts].
Format: Issue; close paraphrase or short quote; exact supplied location; possible relevance; missing facts; verification checklist.
Hard nos: Do not search from memory, invent paragraph numbers, give tax or legal advice, ignore dates, or state a final conclusion.
The blind-paste risk is a real-sounding citation that does not exist or no longer applies.
Then do this: Open the official publication. Verify wording, location, scope, amendments, and effective date.
CLIENT EXPLANATION
Wednesday at 6:45, Owen needs to explain why bookkeeping profit and bank cash differ. His first draft removes an important condition.
COPY THIS
Role: Edit my approved accounting explanation for a non-accountant.
Task: Make it easier to read without changing its meaning.
Details: Audience: [small-business owner]. Approved explanation: [text]. Terms that must stay exact: [accrual, receivable]. Figures allowed: [supplied figures only].
Format: Explanation under 130 words; three short bullets using supplied figures; likely client questions.
Hard nos: Do not invent transactions, promise an outcome, give personal tax advice, remove conditions, or turn “may” into “will.”
The blind-paste risk is simplification that changes the accounting point.
Then do this: Read the edit beside the approved version. Restore every condition and distinction that matters.
FORMULA TEST
Friday at 2:30, Rina gets a duplicate-invoice formula. It looks right. It breaks on blank cells and uses a function her Excel version lacks.
COPY THIS
Role: Review spreadsheet logic as a test designer. The workbook remains the official file.
Task: Explain and challenge the formula I supply.
Details: Platform and version: [Excel 365 or Sheets]. Formula: [formula]. Intended rule: [plain English]. Made-up sample rows: [values]. Edge cases: [blanks, text numbers, duplicates, errors].
Format: Plain explanation; references; failure points; test table with expected results; revised formula labeled UNTESTED.
Hard nos: Do not claim it works, invent ranges, alter live data, or treat sample tests as full assurance.
The blind-paste risk is valid-looking syntax that fails outside one easy example.
Then do this: Test in a copy with normal, blank, boundary, duplicate, error, text-number, and negative cases.
PRIVACY CHECK
Tuesday at 7:18, Cal removes a company name from a ledger export. A memo field still contains an employee’s health detail. The file properties name the client.
Check names, addresses, emails, bank and card numbers, tax IDs, payroll data, health details, customer lists, account descriptions, comments, formulas, links, hidden rows, hidden tabs, file names, and metadata.
Small facts can identify a client when combined. “Store 14, Main Street, Tuesday payroll” may be enough without a name.
Use only tools allowed by firm policy, contract terms, engagement rules, access controls, and retention rules. A delete button in a public chat is not permission to upload a file.
The failure mode is false safety: removing one name while leaving many clues.
Hard nos: No passwords, access codes, bank files, payroll files, restricted evidence, or client documents copied for convenience.
Then do this: If safe redaction ruins the task, use made-up data or ask your firm’s privacy or technology contact.
STUCK
Thursday at 4:58, Ben gets a polished memo with an unexplained number, an unsupported cause, and a conclusion stronger than the file.
COPY THIS
Role: Be a strict evidence editor.
Task: Cut this draft back to what the supplied record supports.
Details: Draft: [redacted draft]. Evidence with source labels: [packet]. Approved conclusion wording, if any: [text or none].
Format: 1) unsupported claims; 2) number-to-source table; 3) contradictions; 4) missing questions; 5) revised draft with [VERIFY] marks.
Hard nos: Do not fill gaps, soften contradictions, invent citations, remove limits, or strengthen the conclusion.
The blind-paste risk is asking for “a better version” and receiving more polish instead of more truth.
Then do this: Resolve every [VERIFY] mark in the actual file. Remove claims you cannot support.
ACCOUNTANT WORKFLOW
Close day at 4:40, Nia wants to paste the first answer into her note. She stops at five gates.
Stop when the task needs missing evidence, restricted information, a tax or legal decision, a final audit conclusion, or a control the model cannot perform.
The failure mode is “looks reasonable” review. Clean writing can hide a weak source trail.
Hard nos: No AI-made evidence. No skipped calculation check. No approval based on tone.
Then do this: Write SOURCE — TRACE — TEST at the top of your next AI-assisted draft.
SOURCES
OpenAI / Chatterji et al. “How People Are Using ChatGPT.” September 15, 2025.
https://openai.com/index/how-people-are-using-chatgpt/
Anthropic. “Be Clear, Direct, and Detailed.” Includes long-document prompt guidance and the reported “up to 30%” test result.
https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct
Anthropic Academy. “The AI Fluency Index.” 2026. Its iteration findings are correlational, not causal.
https://academy.claude.com/tutorials/the-ai-fluency-index
Google Cloud. “Prompt Design Strategies.” Lists objective, persona, constraints, context, and response format.
https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies
These sources support drafting practice. They do not establish accounting accuracy or replace official guidance, firm policy, engagement rules, or professional review.
Hard nos: Do not cite these prompt-craft sources as proof that an accounting conclusion is correct.
Then do this: Verify accounting claims in the official sources required by your firm and engagement.