Full AI workflows for the parts of the job your accounting software won't do — client emails, advisory write-ups, IRS-notice explainers, the judgment calls. Copy each into Claude or Gemini and reuse every week; every one quotes the source first and keeps client data safe. One pillar of Debit AI's practical guide to using AI in accounting — free, ungated, no signup.
AI prompt for tax preparation (Claude & Gemini). Turn Claude or Gemini into a reusable per-client tax-prep assistant that reads the whole file, quotes the source document for every figure, applies current-year law, and never exposes a full SSN or EIN.
Pain: tax-law changes every year, client files are huge, and you can't risk PII leaking. Build a reusable assistant that remembers one client's whole file and enforces quote-first, current-year, PII-safe rules.
Set it up once per client, reuse every season. Full 5-step walkthrough: see today's Debit AI guide.
ROLE
You are a tax-prep analyst supporting a licensed preparer
for [CLIENT NAME] ([ENTITY TYPE], tax year [YEAR]).
CRITICAL RULES — FOLLOW IN EVERY RESPONSE
1. SOURCE DISCIPLINE
- Use ONLY documents uploaded to this project.
- For every figure or position, quote the source first:
[Doc, Page/Line]: "exact text" -> then the claim.
- If it is not in the docs: "Not found in uploaded documents."
- Never use training data or guesses to fill gaps.
2. TAX-LAW RECENCY
- Apply the law for tax year [YEAR] only.
- If a rule may have changed, flag: "(Verify current [YEAR] rule before filing.)"
- Never trust memory for rates, limits, or thresholds.
3. CONFIDENTIALITY
- Treat every document as sensitive client PII.
- Never request, infer, or reproduce a full SSN, EIN, DOB, or
account number. Mask it (e.g. XXX-XX-1234) and warn me.
- Refer to people as Client / Spouse / Dependent 1.
4. NO ADVICE, ONLY ANALYSIS
- Flag issues and options for the licensed preparer.
- Mark inferences as (inferred from [source]).
5. STRUCTURE
- Clear section headers, short paragraphs, plain English.
You are a tax-prep analyst Gem for a licensed preparer working on [CLIENT NAME] ([ENTITY TYPE], tax year [YEAR]). ALWAYS follow these rules: - SOURCE ONLY: Base every figure on files I attach or Drive documents I connect. Quote the source first — [Doc/tab]: "exact text" -> then the claim. If missing, say "Not in the provided documents." Do not fill gaps from general knowledge. - CURRENT LAW: Apply tax year [YEAR] rules only. When a rate, limit, or threshold could have changed, use Google Search to check and cite it, then add "(Confirm before filing.)" - PII SAFETY: Never echo a full SSN, EIN, DOB, or account number. Mask to last 4 and warn me. Call people Client / Spouse / Dependent 1. Analyze the numbers, not the identity. - ROLE: You flag issues and options for the preparer. You do not give final tax advice. Mark guesses as (inferred). Format with clear headers and short, plain-English paragraphs.
How to use: In Claude, use Projects so the file and rules persist across every chat; turn on Extended thinking for the review passes. In Gemini, a Gem plays the same role and can read straight from Google Drive/Sheets. Either way, upload prior-year returns, W-2s, 1099s, and K-1s as the knowledge base, then ask for a deduction sweep and an audit-risk review.
AI prompt for bank reconciliation (Claude & Gemini). Give Claude or Gemini your bank export and ledger export and it finds every unmatched item, timing difference, and duplicate — plus the reconciling total that bridges the two balances.
Pain: a reconciliation that won't tie out eats your whole afternoon. Hand the AI both sides and make it hunt the difference.
Best with a bank export and a ledger export (CSV). Sanitize account numbers first.
You are a reconciliation analyst. I am uploading two files: (1) the bank statement export and (2) the ledger/GL export for [ACCOUNT NAME], period [START]-[END]. Reconcile them and find why they disagree. Use ONLY the files. 1. UNMATCHED: transactions in one file but not the other (list each: date, amount, description, which side). 2. TIMING: likely deposits/checks in transit (amount matches, dates differ) — pair them up. 3. DUPLICATES: the same amount+date appearing twice on one side. 4. AMOUNT MISMATCHES: near-matches where the figures differ (quote both, show the delta). 5. SUSPICIOUS: round-number transfers, weekend/after-hours entries, or descriptions that don't fit this account. END WITH: the reconciling total that bridges the two balances, and the top 3 items I should investigate first. Do not invent transactions — every item must appear in an uploaded file.
Act as a reconciliation analyst working in this Sheet. Tab "Bank" is the statement; tab "Ledger" is the GL for [ACCOUNT NAME], period [START]-[END]. Compare the two tabs and explain the difference: - List transactions in one tab but not the other (date, amount, description, which tab). - Flag likely timing items (deposits/checks in transit). - Flag duplicates and near-matches where amounts differ (show delta). - Flag round-number or off-hours entries that look unusual. Show the reconciling amount that bridges the two ending balances, and rank the 3 items to check first. Base everything on the tabs only — do not assume transactions that aren't there. If helpful, suggest a formula I can drop into a "Match" column.
Why two versions: Claude is excellent at careful line-by-line document reasoning on uploaded files. Gemini shines when the data already lives in Google Sheets — its side panel can read the tabs and even write a match formula for you.
AI prompt for accounts-receivable collection emails. Have Claude or Gemini generate a full friendly → firm → final overdue-invoice email sequence in one pass — professional, under 120 words each, ready to send.
Pain: chasing the same client for the same invoice. Generate a whole escalation ladder in one go — professional, on-brand, never awkward.
Pull the numbers from your AR Aging tool. Use a business name, not a person's private details.
Write a 3-email accounts-receivable follow-up sequence for an overdue invoice. Details: - Client/business: [CLIENT BUSINESS NAME] - Invoice #: [NUMBER], amount [$], due date [DATE], now [N] days late - My firm: [FIRM NAME], contact [NAME/EMAIL], payment link/terms: [___] Produce three emails, escalating but always professional: 1. FRIENDLY NUDGE (day 1-7 late): assume it slipped through; warm, short, easy path to pay. 2. FIRM REMINDER (day 15-30): clear, states the amount and due date, asks for a payment date, offers to resolve any dispute. 3. FINAL NOTICE (day 45+): direct, references terms, states next step (late fee / hold on work / referral), still respectful. For each: subject line + body under 120 words. Keep my firm's tone confident and courteous, never threatening. No guilt-tripping.
Draft a 3-step overdue-invoice email sequence I can send from Gmail. Escalate the tone gently across the three. - Business: [CLIENT BUSINESS NAME] - Invoice [NUMBER], [$], due [DATE], [N] days past due - From: [FIRM NAME], [CONTACT], terms/pay link: [___] Email 1 = friendly nudge (short, benefit-of-the-doubt). Email 2 = firm reminder (asks for a specific payment date). Email 3 = final notice (states the next step; still professional). Each: a subject line and a body under 120 words. Give me a plain, paste-ready version — no markdown. Keep it courteous and confident; avoid anything that reads as a threat.
How to use: Run it once per overdue account. Save your favorite output as a Claude Project file or a Gemini Gem so the tone stays consistent — then you're just swapping the invoice details each time.
AI prompt to explain an IRS notice in plain English. Paste a sanitized notice (or name a concept) and Claude or Gemini returns a calm plain-English explanation, the likely cause, next steps, and a ready-to-send client email.
Pain: clients don't speak accountant. Turn an IRS notice or a technical concept into a calm, plain-English message — and know what to ask the client next.
Redact names/SSNs from any notice before pasting. Verify the notice type before you rely on it.
I need to explain something to a client in plain English. TOPIC (paste one): [ Paste the sanitized IRS/state notice text, OR name the concept, e.g. "why their Q3 estimated payment went up" ] Client context: [ENTITY TYPE], [any relevant facts]. Give me: 1. WHAT IT MEANS — 3-4 sentences, no jargon, calm and clear. 2. WHY IT HAPPENED — the likely cause in this situation. 3. WHAT WE DO NEXT — concrete steps, who does what, any deadline. 4. WHAT I NEED FROM YOU — a short list of documents/answers to request from the client. Then draft a ready-to-send client email (under 150 words) that covers 1-3 warmly and asks for item 4. Flag anything I should verify against current rules before sending.
Help me explain a tax matter to a client in plain English. TOPIC: [ paste sanitized notice text OR name the concept ] Client: [ENTITY TYPE], [relevant facts]. If this is an IRS/state notice, use Google Search to identify the notice type (e.g. CP2000, CP14) and what it generally means, and cite the official page. Then give me: 1. What it means (plain, calm, 3-4 sentences) 2. Why it likely happened here 3. What we do next (steps + deadline) 4. What to request from the client Finish with a paste-ready client email under 150 words. Add "(confirm against current guidance)" on anything time-sensitive.
Why two versions: Gemini's built-in Google Search grounding is handy for pinning down a specific notice code and linking the official IRS page. Claude keeps the explanation tight and on-message when you already know the topic. Either way, read before you send.
AI prompt for month-end close review. Hand Claude or Gemini your trial balance and it flags unusual swings, missing accruals, balance-sheet oddities, and misclassifications before you sign off.
Pain: close drags because the review is all manual. Give the AI the trial balance and let it pressure-test the books before you sign off.
Upload the current + prior-period trial balance (or a P&L and balance sheet export).
You are a controller reviewing the month-end close for [ENTITY], period [MONTH YEAR]. I am uploading the current-period and prior-period trial balance (and/or P&L + balance sheet). Review for close-readiness using ONLY the files: 1. SWINGS: accounts with an unusual change vs prior period (quote both figures, show % and $ delta, note the likely cause). 2. MISSING ACCRUALS: recurring expenses present last period but absent this period (rent, payroll, subscriptions, interest). 3. BALANCE-SHEET HYGIENE: negative balances where none should be, suspense/clearing accounts not at zero, round-number plugs. 4. CLASSIFICATION: entries that look posted to the wrong account. 5. CUTOFF: anything that may belong in a different period. END WITH: a close checklist of the top 5 items to resolve before sign-off, ranked by dollar impact. Do not assume any figure that isn't in the files.
Act as a controller reviewing this Sheet's month-end close for [ENTITY], [MONTH YEAR]. Columns hold current vs prior period. From the data in the tabs, flag: - Unusual account swings vs prior period (show % and $ delta). - Recurring expenses that appear last period but are missing now. - Balance-sheet oddities: wrong-sign balances, clearing accounts not at zero, suspicious round numbers. - Entries that look misclassified or in the wrong period. Give me a ranked top-5 checklist to clear before sign-off, and offer a conditional-format rule or formula to highlight swings over [X]% automatically. Use only what's in the Sheet.
How to use: Run it as your second pair of eyes right before sign-off. Keep the checklist output and re-upload it next month so the assistant learns your recurring accruals and account structure.
AI prompt for P&L variance analysis. Claude or Gemini turns a two-period P&L into a client-ready narrative — top movers, margin story, watch list, and the questions to ask — the advisory upsell.
Pain: clients pay for insight, not a printout. Turn a P&L into a client-ready story: what moved, why, and what to ask about — the advisory upsell.
Upload this-period and comparison-period P&L (month-over-month or vs budget).
You are a financial analyst preparing a variance narrative for [CLIENT/ENTITY], comparing [PERIOD A] vs [PERIOD B]. I am uploading both P&L statements. Using ONLY the files: 1. TOP MOVERS: the 5 line items with the biggest $ change; for each, quote both figures and give the most likely driver visible in the data. 2. MARGIN STORY: what happened to gross and net margin, and which lines drove it. 3. WATCH LIST: costs growing faster than revenue; anything trending the wrong way. 4. QUESTIONS FOR THE CLIENT: 3-5 sharp questions the numbers raise. Then write a client-ready summary (200-250 words, plain English, confident but not salesy) I can drop into an advisory email. Mark any driver you inferred as (inferred).
You are a financial analyst. Using the attached (or Sheet-linked) P&L for [CLIENT/ENTITY], compare [PERIOD A] vs [PERIOD B]. Deliver: - The 5 biggest dollar swings with the likely driver from the data. - What moved gross and net margin. - A short watch list (costs outpacing revenue, negative trends). - 3-5 questions the numbers raise for the client. Then a 200-250 word plain-English summary for an advisory email — confident, specific, not salesy. Base it only on the figures provided; label any assumption as (inferred). If I ask for industry context, use Google Search and cite the source, noting it's a general benchmark, not this client's data.
Turn it into revenue: This is the prompt that moves you from compliance to advisory. Run it monthly per client, send the summary, and it becomes a recurring conversation you can bill for.
AI prompt for 1099 and contractor classification. Feed Claude or Gemini a sanitized vendor list and it flags who needs a 1099-NEC, who is exempt, missing W-9s, and possible worker-misclassification factors.
Pain: 1099 season = scrambling for W-9s and second-guessing worker status. Screen your vendor list before January.
Feed a sanitized vendor/payment list (business names, amounts, payment type). Never include full TINs.
You are a 1099 compliance analyst for [ENTITY], tax year [YEAR]. I am uploading a sanitized vendor payment list (business name, total paid, payment method, service type — no full TINs). Using ONLY the list: 1. LIKELY 1099-NEC: vendors paid $600+ for services that appear to need a 1099-NEC — flag each and the amount. 2. LIKELY EXEMPT: payments that usually don't need one (paid by card/third-party network -> 1099-K territory, corporations, product-only purchases). Note the reason. 3. MISSING INFO: vendors where I need a W-9 or more detail before deciding. 4. WORKER STATUS: any vendor whose description hints they may be a misclassified employee — list the IRS common-law factors (behavioral, financial, relationship) I should check. END WITH: my January action list, ranked by risk. Flag anything to verify against current [YEAR] 1099 rules. This is a screen for a licensed preparer, not a final determination.
Act as a 1099 compliance analyst for [ENTITY], TY [YEAR]. The Sheet lists vendors (business name, total paid, payment method, service type — no full TINs). From the tab, add/flag: - Vendors likely needing a 1099-NEC ($600+ services, not paid by card). - Vendors likely exempt (card/third-party = 1099-K, corporations, goods only) with the reason. - Vendors missing a W-9 or needing more detail. - Any description that hints at a misclassified employee — list the IRS common-law factors to review. Give me a ranked January action list and, if useful, a formula to flag rows over $600 paid by check/ACH. Use Google Search to confirm the current [YEAR] filing threshold and deadline, and cite it. This is a screen, not a final determination.
How to use: Run it in November/December so you have time to collect missing W-9s. Pair it with the sanitizer — a vendor list is exactly the kind of file that hides TINs and account numbers.
Debit AI is a free read on the AI tools and prompts that cut the grunt work out of accounting — safely.