::play 01 · Variance Commentary First Draft
Claude (enterprise tier) or local Ollama if the data is sensitive. Use anonymized account names if cloud.
Turn a numeric variance table into board-ready narrative commentary in minutes, not hours, while keeping every number recomputed and every claim traceable.
You are an FP&A analyst drafting board commentary. I will give you a variance table showing actual vs. budget for the current month and YTD across [REVENUE / COGS / OPEX categories]. For each variance greater than [$THRESHOLD] or [%THRESHOLD], draft 2-3 sentences of commentary that: (1) states the variance in dollars and percent, (2) proposes the most likely driver based on the account name and the direction of the variance, (3) flags any variance you cannot explain from the data alone so I can investigate. DO NOT speculate beyond what the data supports. DO NOT round or alter the numbers I gave you. End with a list of 'questions to ask operations' for any variance you flagged. Here is the data: [PASTE ANONYMIZED VARIANCE TABLE]. Output as a markdown table: Account | Variance $ | Variance % | Proposed Driver | Confidence (H/M/L) | Investigate?
::what to notice · How often the AI proposes a plausible-sounding driver that does not match what actually happened. The 'Confidence' column is your tell. Anything Medium or Low gets a human investigation before it appears in the deck.
::trap · AI loves to confidently explain variances using generic finance vocabulary ('higher marketing spend driving revenue growth') that is technically possible but factually wrong. Never ship variance commentary without a human verifying the proposed driver against actual operational data.
::play 02 · Board Deck Narrative Synthesis
Claude with long-context window. Anthropic's long-context performance is materially better for this specific task than alternatives as of 2026.
Synthesize 40+ pages of monthly close materials into a 3-page executive summary that the CEO and board can read in five minutes.
You are preparing the CFO's three-page narrative for the board. I will paste the full monthly close package: variance commentary, KPI dashboard, cash flow summary, operational updates, and risk register. Produce a three-page narrative with these sections: (1) Headline — the one number and one trend the board must understand this month, in two sentences. (2) What Worked — three specific wins with the actual numbers. (3) What Didn't — three specific misses with the actual numbers and what we're doing about each. (4) Cash & Liquidity — current cash position, 13-week forecast, covenant headroom. (5) Forward Look — top three risks and top three opportunities for the next 90 days. Use only numbers and facts present in the materials I gave you. Where you are uncertain, write [VERIFY] inline. Do NOT invent metrics. Do NOT smooth bad news. Tone: direct, audit-trail-grade, no hype. Length: 1,200 words maximum. Source materials: [PASTE]
::what to notice · Where the AI smooths bad news. Models are trained to be agreeable; that is the opposite of what a board narrative needs. Every [VERIFY] tag is a number you personally re-source from the close package.
::trap · Models will compress and lose the negative. A 7% revenue miss in one line item gets buried under 'overall performance in line with plan.' Read the AI draft against the worst line in the original materials. If the worst line is missing from the narrative, the narrative is wrong.
::play 03 · Accounting Memo Drafting (ASC Research)
Claude or ChatGPT Enterprise. NEVER use AI as the authoritative source for the citation itself — verify against the actual Codification.
Draft a technical accounting memo on a specific transaction or policy question with proper ASC citation structure, ready for controller review.
Act as a senior technical accounting manager drafting a memo for the controller's file. Topic: [TRANSACTION OR POLICY QUESTION]. Facts: [PASTE FACT PATTERN — anonymize counterparty if material non-public]. Required structure: (1) Background and facts (1 paragraph). (2) Question presented (1 sentence). (3) Applicable guidance — list the specific ASC topics and subtopics you believe apply, with paragraph references. Mark each citation [VERIFY AGAINST CODIFICATION]. (4) Analysis — apply the guidance to the facts. (5) Conclusion — state the accounting treatment. (6) Journal entry illustration. (7) Disclosure considerations. Be conservative. If two interpretations exist, present both. If guidance is genuinely ambiguous, say so and recommend consultation with the audit firm. Do NOT invent ASC paragraph numbers — if you are not certain of the exact paragraph, write 'See ASC [Topic] generally' and let the reviewer pinpoint.
::what to notice · Every single ASC citation. Models hallucinate paragraph numbers routinely. The memo is unusable until a human verifies each citation against the FASB Codification (or IFRS Standards) and corrects any that are wrong.
::trap · The biggest finance-AI scandal of 2025 was a Big 4 firm shipping a client memo with hallucinated ASC 606 paragraph references. The firm rebuilt the memo, ate the fees, and added a mandatory verification step. Do not be that controller.
::play 04 · Vendor Contract Clause Check
Claude — best at long-document reasoning. Acceptable to use cloud for vendor contracts that do not contain customer PII. Use local Ollama for anything sensitive.
Pre-screen a vendor or customer contract for finance-relevant clauses (revenue recognition triggers, payment terms, auto-renewal, indemnification caps, termination) before sending to legal.
You are a senior finance reviewer scanning a contract for finance-impact clauses. Read the attached agreement and produce a structured summary covering: (1) Contract term, auto-renewal, and termination rights — quote the exact language. (2) Payment terms, late fees, and any variable consideration. (3) Performance obligations and acceptance criteria (relevant to ASC 606 revenue recognition). (4) Refund, credit, or service-level penalty provisions. (5) Indemnification caps and limitations of liability. (6) Change-of-control or assignment provisions. (7) Any 'most favored nation,' exclusivity, or non-compete clauses. (8) Any clauses that look unusual, asymmetric, or expensive. For each finding, quote the exact contract language and cite the section number. Do NOT paraphrase the legal language — quote it verbatim so the controller and legal can verify. End with a 'Top 5 issues to escalate' list. Contract: [PASTE]
::what to notice · How well the AI quotes verbatim vs. paraphrasing. Paraphrased contract language is useless to legal. Insist on direct quotes with section references so verification is mechanical.
::trap · AI will confidently summarize a clause in a way that subtly misstates the obligation. The risk is not that you'll miss something — it's that you'll trust the summary and never read the actual clause. Always read the quoted text yourself before the contract goes to signing.
::play 05 · Cash Flow Forecast Driver Review
Claude or ChatGPT Enterprise. Use anonymized aggregate cash data; never paste individual customer payment timing.
Pressure-test a 13-week cash forecast by having AI propose the questions a treasurer or CFO would ask, before the actual treasurer or CFO asks them.
Act as a skeptical CFO reviewing a 13-week cash forecast prepared by a treasury analyst. I will give you the forecast assumptions: receipts by category, disbursements by category, opening balance, and any known one-time items. Your job is to: (1) Identify any assumption that seems aggressive or unsupported. (2) Identify any disbursement category that looks low based on historical patterns I provide. (3) Propose three downside scenarios — payroll timing shifts, a top-3 customer pays 15 days late, an unexpected tax payment. (4) Calculate the minimum cash balance under each scenario and flag any week where we breach our minimum operating cash threshold of [$X]. (5) List the five questions you would ask the analyst before approving the forecast. Do NOT do the math yourself — show the formula and let me recompute in Excel. Forecast and history: [PASTE]
::what to notice · The skepticism quality. A good prompt produces questions like 'why did Q3 historical disbursements jump 18% and is that reflected in the forecast?' A bad prompt produces generic CFO-speak. Tune the prompt with more historical context if the questions are too vague.
::trap · Never let the AI do the actual cash math. Always recompute the minimum balance and the scenario impacts in Excel or Python. AI arithmetic on 90 cells of forecast data is unreliable enough that treasury professionals have signed off on impossible cash positions because they trusted the AI summary.
::play 06 · Audit Prep Documentation Pattern (PBC List)
Local Ollama strongly preferred — audit prep documents often contain customer-level detail. Cloud only if every data element is genuinely anonymized.
Build the documentation package an external auditor needs for a specific area (revenue, leases, inventory, comp) so the prepared-by-client list comes back faster and cleaner.
Act as a senior accountant preparing audit documentation for [AUDIT AREA: e.g., revenue recognition under ASC 606]. The auditor will request: process narratives, walkthrough documentation, controls testing evidence, journal entry sampling support, reconciliations, and significant estimates documentation. Produce a checklist of every document I should prepare in advance, organized by: (1) Standard PBC items the auditor will definitely request. (2) Likely follow-up items based on common audit findings in this area. (3) Areas where the auditor will likely focus given recent PCAOB inspection findings or AICPA hot topics. For each item, specify: document name, owner, source system, expected format, and any cross-references. End with a 'top three risk areas' summary the controller should personally review before the auditor's first meeting. Be specific to [INDUSTRY] and [COMPANY SIZE / FILING STATUS].
::what to notice · Whether the AI knows current PCAOB inspection focus areas. As of 2026 these include revenue cutoff, internal controls over journal entries, and going-concern evaluation. If the output ignores these, the prompt needs the year and the audit firm tier as additional context.
::trap · Do not let the checklist replace your professional judgment about your own company's risk areas. The AI's 'standard PBC list' is a starting point. The actual audit risk is in the line items that are unique to your business — those won't be on any standard list.
::play 07 · SEC Filing / Regulatory Document Review (Draft Review Only)
Claude with long-context. Cloud is acceptable for already-public-equivalent drafts. NEVER paste pre-release MNPI for unfiled documents into a non-enterprise AI.
Pre-screen an SEC filing draft, proxy, 10-K MD&A, or regulatory response for internal inconsistencies, unsupported claims, missing disclosures, and tone problems before the final round of review.
Act as a senior SEC reporting specialist reviewing a draft [10-K / 10-Q / proxy / S-1] before it goes to the audit committee. Your job is to flag — not fix — issues. Produce a structured review covering: (1) Internal consistency — any number, percentage, or trend that appears in two places and does not match. Cite the page and section of each occurrence. (2) Unsupported claims — any forward-looking statement, market position claim, or competitive assertion that does not have a clear basis in the filing's data. (3) Missing standard disclosures — anything required by Reg S-K or Reg S-X that appears absent or thin. (4) MD&A tone — anything that reads like marketing rather than analysis. (5) Risk factor staleness — anything that looks copy-pasted from prior years and may not reflect current risk. (6) Cross-reference integrity — any reference to an exhibit, schedule, or section that may be broken. DO NOT propose actual language fixes — that is for the controller and outside counsel. Just flag. Draft: [PASTE]
::what to notice · False positives. AI tends to over-flag in regulatory filings because it doesn't know the prior-year baseline. Triage findings against the prior filing before escalating to the audit committee.
::trap · AI cannot make a materiality judgment for you. Every flagged item still goes through your firm's materiality framework before becoming a real change. Do not let the AI's volume of findings drive the disclosure decision.
::play 08 · Tax Research First Pass (Circular 230 Constraint)
Claude or ChatGPT Enterprise. Use anonymized fact patterns. NEVER paste actual taxpayer-identifying information.
Build a first-draft tax research memo on a Code section, regulation, or recent guidance — knowing that every citation and every conclusion will be independently verified by the licensed preparer.
Act as a tax research associate preparing a memo for a licensed CPA/EA who will independently verify every citation. Question: [TAX QUESTION]. Anonymized facts: [PASTE]. Produce a research memo with: (1) Question presented. (2) Brief facts. (3) Applicable Code sections, Treasury Regulations, and recent guidance — give the citation in the format §XXX(x)(x), Reg. §X.XXXX-X, Rev. Rul. XXXX-XX. Mark every citation [VERIFY]. (4) Analysis applying the authority to the facts. (5) Conclusion. (6) Alternative positions and their relative authority. (7) Practical considerations including penalty exposure under §6662 if the position is challenged. Be conservative. Flag any position that would require disclosure on Form 8275. Do NOT cite cases or rulings unless you are confident they exist — if uncertain, say 'There may be relevant case law in [AREA]; researcher should verify in CCH/RIA/Bloomberg Tax.'
::what to notice · How many citations turn out to exist when you check them in the actual tax research database. The fail rate is non-trivial. Treat the memo as a starting outline that gets rebuilt against verified authority before any client work product is delivered.
::trap · IRS Circular 230 §10.34 requires the preparer to have a reasonable basis for every position. 'I asked an AI' is not a reasonable basis. Every Code citation, Reg citation, and case citation in the final memo must be independently verified in a paid tax research service. The preparer signs the return, not the model.
