A junior on a close team pastes a trial balance into ChatGPT, asks “explain the variances,” and gets back a confident paragraph citing “ASC 842-17.” It reads clean enough to forward to a partner. The cited section does not exist. Same junior, same trial balance, a four-line scaffold pasted above the data — and the model returns a flux memo with every line over materiality quantified and categorized, ready for review in two minutes.
The scaffold did the work.
This guide gives you the scaffold, paste-ready prompts for the four workflows that eat the most hours — close, tax memos, bookkeeping, audit — and the guardrails that keep you out of trouble.
CARV: the only scaffold you need
Context, Ask, Rules, Verify. Four slots. Paste this template above any accounting prompt and the output becomes usable on the first try instead of the fifth:
Context: You are [role] for [entity], period [YYYY-MM]. Materiality: $[X]. Source data is pasted below.
Ask: [single concrete deliverable].
Rules:
- Output format: [table | memo | bullet list | Excel formula].
- Tone: [neutral professional | client-facing plain English | review-note shorthand].
- If you do not have authority or data to support a claim, write "INSUFFICIENT — need: [X]" and stop.
Verify: Before returning, check that [2–3 specific checks].
Data:
[paste TB / GL / CSV / authority text]Verify carries the weight. It forces the model to self-check before it answers — and it tells you exactly which clauses you must reconfirm before sign-off. Treat each Verify line as dual-purpose: the model attempts it, you reconfirm anything that could land on a workpaper. The governing rule:
If the model cannot cite the source or tie out to the total, it must refuse — not guess.
Workflow 1: Close and flux analysis
Most of the close week disappears into flux commentary, and the outlier you missed is the one the controller circles.
Flux commentary prompt (paste into ChatGPT or Claude):
Context: You are senior accountant for ACME Industrial, period 2026-04. Materiality: $25,000. Source is the P&L variance table below.
Ask: Draft flux commentary for every line where |variance| ≥ materiality.
Rules:
- Output format: 2-column table — Account | Commentary.
- Tone: neutral professional, one sentence per account.
- Categorize each variance as one of: volume | price | timing | one-time | mix | unknown.
- If a category is unclear, write "unknown — request driver from FP&A."
Verify: (1) Every flagged row appears in the output. (2) Sum of variances reconciles to total variance. (3) No line uses two categories.
Data:
[paste variance table]For reconciliations, the same shape with a different Ask:
Context: You are a corporate accountant reviewing the [Account] reconciliation for [Entity], [YYYY-MM]. Materiality: $[X].
Ask: Group reconciling items into three buckets and rank by suspicion.
Rules:
- Buckets: (a) timing, expected to clear (b) unexplained, needs action (c) reclass.
- Output: table — Item | Bucket | Suspicion (low/med/high) | Suggested next step.
Verify: (1) Every item assigned exactly one bucket. (2) Items total to reconciliation difference. (3) Each "high" suspicion item names the test to run next.
Data:
[paste reconciliation lines]Copilot in Excel — in-cell variance flagger. Excel Copilot is range-aware, which is the entire reason to use it here instead of pasting numbers into a chat window. Select the variance column, then:
Compare values in the selected range against the prior-period column to its left. Return a new column "Flag" with "REVIEW" if |variance| ≥ 25000 and "OK" otherwise. Add a "Direction" column with "↑" if current > prior, "↓" if lower. Do not modify other columns.A 200-line P&L typically collapses to 8 to 12 rows worth reading.
Workflow 2: Tax research and client memos
Hallucinated citations are the failure mode here. A generic prompt will fabricate a Revenue Ruling and stand behind it. Lock that down inside Rules: primary authority only, refusal otherwise. Claude holds long authority text more cleanly, which matters here.
Authority-anchored memo prompt:
Context: You are a US tax associate. Client: [name], entity: [LLC/S-corp/C-corp], tax year: [YYYY].
Ask: Write a tax memo on this issue: [issue].
Rules:
- Cite only primary authority: IRC, Treasury Regulations, Revenue Rulings, Revenue Procedures, Tax Court decisions, IRS Notices.
- If primary authority does not directly resolve a sub-issue, write "Authority does not clearly address this; flag for partner review" and stop — do not invent or paraphrase secondary sources.
- Output sections: Facts | Issue | Authority | Analysis | Conclusion.
- Length: 600 words max.
Verify: (1) Every cited authority is real and active (no superseded rulings). (2) Each authority quoted ≤ 25 words. (3) Conclusion answers the Ask in one sentence.
Issue and facts:
[paste]You still check citations yourself. The point is that the model refuses on the sub-issues where it would otherwise improvise. Run the issue once without the Rules slot and you will see a fabricated Rev. Proc. inside a paragraph.
Plain-English client email prompt (run after the memo lands):
Context: Same client and tax year. Convert the attached memo into a client email.
Ask: One-page client email at an 8th-grade reading level.
Rules:
- No citations in the body. Move citations to a single "Sources" line at the bottom.
- Open with the answer in the first sentence.
- Translate every technical term into plain English the first time it appears.
Verify: (1) First sentence is the answer. (2) No bullets longer than one line. (3) Email ends with one clear next step for the client.
Memo:
[paste prior output]Clients read sentence one and skim the rest, which is why the answer-first opener carries the email. ChatGPT handles this translation pass cleanly.
Workflow 3: Bookkeeping and transaction categorization
The trap with bulk categorization is losing the audit trail or coding into accounts that do not exist in the chart. The prompt below enforces both — full table plus a separate exception list, with a refusal token when no account fits.
Context: You are a bookkeeper for [Entity]. Chart of accounts pasted below. Materiality for review: $[X].
Ask: Categorize each transaction in the CSV. Return two outputs.
Rules:
- Output 1: full table — Date | Vendor | Memo | Amount | Account (must be a row in the COA) | Confidence (high/med/low).
- Output 2: exception list — every transaction with Confidence ≠ high, amount ≥ materiality, ambiguous vendor name, or new vendor not previously seen.
- Never invent an account name not in the COA. If none fits, mark account as "ASK_BOOKKEEPER".
Verify: (1) Every transaction appears once in Output 1. (2) Every "ASK_BOOKKEEPER" appears in Output 2. (3) Sum of categorized amounts equals total of the CSV.
COA:
[paste]
CSV:
[paste]Output 1 is the receipt for the audit trail. Output 2 is what you actually work from.
Workflow 4: Audit workpapers and review notes
Audit work has the heaviest review burden and the highest confidentiality exposure. Two prompts cover most of the daily load.
Tickmarks → procedures narrative:
Context: You are an audit senior on [Client], FY [YYYY]. Account: [name]. Population total: $[X]. Sample size: [n]. Sampling method: [method]. Tickmark legend below.
Ask: Convert the raw tickmark notes into a procedures narrative for the workpaper.
Rules:
- One paragraph per procedure performed.
- Each paragraph: (a) procedure performed, (b) population/sample reference, (c) result, (d) conclusion.
- Use only information present in the tickmarks. If a tickmark is unclear, write "[CLARIFY: <code>]".
Verify: (1) Every tickmark is reflected in the narrative. (2) Sample size and method match the workpaper header. (3) Conclusion matches firm review standards.
Tickmarks:
[paste]Reviewer-ready responses to open review notes:
Context: Responding to partner/manager review notes on [Workpaper ID].
Ask: Draft a response to each note that closes it.
Rules:
- One response per note, in the firm's review-note shorthand.
- Each response: action taken | reference to updated workpaper section | sign-off initials placeholder.
- If a note requests something not yet performed, draft both the planned procedure and the note response.
Verify: (1) Every note has a response. (2) No response references a workpaper section absent from the input. (3) No response says "noted" alone.
Notes:
[paste]Review notes compound. The second prompt unsticks the workpaper before it freezes the close.
The three failures that get accountants in trouble
1. The phantom ASC. The model fabricates an accounting standard, a Revenue Ruling, or a PCAOB paragraph and writes about it with full confidence. The fix is the Rules slot: require primary authority, force “INSUFFICIENT” or “flag for partner review” when the model would otherwise improvise.
2. The leaky paste. Client names, SSNs, payroll detail pasted into a consumer chat window train someone’s model or sit in a log you do not control. The first line of defense is redacting in Excel or Word before anything reaches the chat window — find-and-replace, then paste. The second is staying inside the enterprise tier of ChatGPT or Claude with training opt-out enabled. As a session-level guard inside that enterprise tier, you can start with:
Before processing, replace every occurrence of the following with placeholders in your response:
- Client legal name → CLIENT_A
- Entity EIN / SSN / TIN → [TAX_ID]
- Bank account / routing numbers → [ACCOUNT_X]
- Employee names → EMP_<n>
Return only the redacted version. Wait for further instruction.That redacts the output. It does not undo a paste that already hit the wire — pre-redact in the file. Never paste payroll detail, audit client identifiers, or unredacted general ledger extracts into a consumer chat. Use Copilot in Excel for anything that has to stay inside the workbook.
3. The unreviewed handoff. Signing your name to a memo, a workpaper, or a reconciliation that you only skimmed. The model is a first-year staff. The sign-off is still yours. Tie every number to source before your initials touch it.
Tool defaults for 2026
Three rules, each anchored to one concrete task:
- Claude for tax memos and any prompt with more than five pages of authority text — the longer the context, the cleaner Claude tends to hold the thread.
- ChatGPT for client-email translation and ad-hoc flux drafts where the prompt is short and the output is conversational.
- Copilot in Excel for variance flagging, sampling work, and anything that has to operate on a live range without re-uploading data.
Most weeks you will hand work off mid-flow: the memo starts in Claude and the client email translates in ChatGPT, or variance flagging stays in Excel and the commentary moves to chat. Store the eight prompts in a shared library — a firm prompt folder, a Notion page, even a pinned Teams post — so the team uses the same versions.
Your one-page operating kit
Scaffold. Paste CARV above every prompt:
Context | Ask | Rules | Verify | DataVerify rule. If the model cannot cite the source or tie out to the total, it must refuse — not guess.
Prompts you now have.
- Flux commentary (close).
- Reconciliation suspicion-sort (close).
- Copilot in-cell variance flagger (close).
- Authority-anchored tax memo (tax).
- Plain-English client email (tax).
- Bulk categorization with exception list (bookkeeping).
- Tickmarks → procedures narrative (audit).
- Reviewer-ready review-note responses (audit).
Guardrails. Phantom ASC, leaky paste, unreviewed handoff. The redaction snippet sits above every client data block in a consumer tool.
You have the prompts. You have the scaffold. You have the guardrails. No CPE. No new subscription. Open it tomorrow and use it.