The Prompt Library Every CFO Should Steal

TL;DR. Ten prompts you can paste into Claude or GPT-5 today. Grouped by function: variance (3), cash flow (2), reporting (2), scenario modeling (2), and interview prep (1). Each one has a template, a target LLM, and a note on where it goes wrong. Copy the whole page into a Claude Project and you have a working CFO prompt library by lunch.

Every CFO who uses AI seriously ends up building a private prompt library. The problem is you spend six weeks reinventing prompts other people already got right. This is the shortcut. Ten prompts I actually run, tested against real financials over the last six months, grouped so you can lift only the ones you need.

Rules of the road before you use them. Replace bracketed values with your data. Never paste unredacted client-identifiable numbers into a public consumer chatbot; scale or synthetic your inputs first. And keep a “prompts.md” file in your finance shared drive so the whole team runs the same versions.

Variance Analysis (Prompts 1 to 3)

Prompt 1. Weekly Variance Narrative

What it does. Turns a weekly actual-vs-budget-vs-prior-week P&L into a five-sentence variance narrative fit for a board email.

You are a CFO writing a weekly variance narrative for the executive team.

Input (CSV): [PASTE P&L WITH COLUMNS: Line Item, Actual, Budget, Prior Week, Variance $, Variance %]

Rules:
1. Lead with the 3 largest absolute-dollar variances.
2. For each, propose 2 plausible drivers grounded in the line items I gave you. Do not invent.
3. Flag any line over 15% versus budget OR over 10% versus prior week.
4. End with one sentence answering "so what" for the CEO.
5. Plain prose. No bullet lists over 5 items.

Business context: [ONE SENTENCE: sector, ARR, gross margin]
Week ending: [DATE]

Which LLM. Claude Sonnet 4.5. Grounds its answer in your rows instead of pattern-matching to generic finance prose.

Common failure. Missing column headers. The model then invents a comparison that isn’t there.

Prompt 2. Month-End Budget-to-Actual Walk

What it does. Produces a bridge narrative from budgeted EBITDA to actual EBITDA, quantifying each driver.

Act as an FP&A analyst. Build a budget-to-actual EBITDA walk for [MONTH].

Budgeted EBITDA: $[X]
Actual EBITDA: $[Y]

Line-item bridge inputs (each row = Line, Budget, Actual, Delta):
[PASTE ROWS]

Produce:
1. A narrative walk from budget to actual, hitting the 5 largest deltas in order of size.
2. Each step names the driver in one sentence.
3. End with the top 2 drivers to watch next month, with the leading indicator I should track for each.

Ignore deltas under $[THRESHOLD] as noise.

Which LLM. Claude Sonnet or GPT-5, both fine.

Common failure. Not setting a materiality threshold, so the model treats every $1,200 travel line as important.

Prompt 3. Anomaly Sweep Across the Trial Balance

What it does. Scans a trial balance snapshot for accounts moving unusually versus a trailing average. Catches booking mistakes before they hit the P&L.

You are an accountant doing pre-close anomaly detection.

Data (CSV): Account, Current Balance, 3-Month Rolling Avg, 6-Month Rolling Avg

Rules:
1. Flag any account where current balance is 3x or more the 3-month rolling average AND absolute delta is over $[THRESHOLD].
2. For each flag, hypothesize whether the driver is a booking error, an accrual timing issue, or a real change in operations.
3. Skip accounts with 6-month rolling avg below $[SMALL DOLLAR NOISE FLOOR].
4. If nothing flags, say so.

Output: list of flags only, each with account name, current balance, 3-month avg, and one-line hypothesis.

Which LLM. Claude with the analysis tool turned on, or GPT-5 with data analysis on. You want actual math, not estimation.

Common failure. Skipping the noise floor. The model then flags 40 tiny accounts and buries the two real ones.

Cash Flow (Prompts 4 and 5)

Prompt 4. 4-Week Rolling Cash Outlook

What it does. Combines current cash, last 8 weeks of net cash flow, and known items to produce a 4-week direct forecast plus a tight-week callout.

Act as a CFO running a 4-week direct cash forecast.

Current cash: $[X] as of [DATE]
Minimum operating cash: $[MIN]
Last 8 weeks net cash flow (weekly, oldest first): [W1, W2, ..., W8]
Known upcoming items:
- Payroll: [DATES, AMOUNTS]
- AP runs: [DATES, AMOUNTS]
- Debt service: [DATE, AMOUNT]
- Large customer receipts: [DATE, AMOUNT, PROBABILITY]

Produce:
1. Week-ending cash table for W+1 through W+4 (base case only).
2. One paragraph naming the tight week and the driver.
3. One line on the single biggest source of uncertainty.

No preamble.

Which LLM. Claude or GPT-5. Claude tolerates messy inputs slightly better.

Common failure. Not giving the model a minimum operating cash floor. Without it, “the forecast looks fine” is meaningless.

Prompt 5. AR Aging Collection Priority

What it does. Takes an AR aging report and produces a prioritized call list for collections, ranked by expected cash impact and win probability.

You are an AR manager building a collection call list.

Aging data (CSV): Customer, Total Outstanding, Current, 1-30, 31-60, 61-90, 90+, Last Payment Date, Notes

Rules:
1. Rank customers by expected collectible dollars in the next 14 days (Total Outstanding x collection probability).
2. Collection probability: 90% for current, 70% for 1-30, 40% for 31-60, 20% for 61-90, 5% for 90+.
3. Produce top 10 as a table: Customer, Outstanding, Bucket Mix, Expected Cash, Suggested Action.
4. Suggested Action options: friendly reminder, firm call, credit hold, write off, escalate.

Order table by Expected Cash descending.

Which LLM. Either.

Common failure. Feeding a raw aging with customer PII intact. Redact names first or run on the enterprise tier.

Reporting and Executive Summaries (Prompts 6 and 7)

Prompt 6. Board Deck Bullet Points from a P&L

What it does. Turns your monthly close file into a five-slide board deck outline with speaker notes for each.

Draft a 5-slide board-meeting financial update for [MONTH].

Inputs:
- P&L (actual vs budget vs prior year): [PASTE]
- Cash position: [$X, with runway in months at current burn]
- Top 3 wins this month: [LIST]
- Top 3 risks this month: [LIST]

Slide structure:
1. Financial headline (1 chart, 3 bullets)
2. Revenue detail (drivers of the quarter)
3. Margin and cost (biggest movers)
4. Cash and balance sheet
5. Wins, risks, and asks for the board

For each slide give: title, 3 bullets max, and 2 sentences of speaker notes. No hedging language. If you need a number I did not give you, put "[NEED: X]" so I can fill it in.

Which LLM. GPT-5. Its default prose voice for boardroom audiences is tighter than Claude’s out of the box.

Common failure. Not asking for “[NEED: X]” placeholders. The model then quietly makes up a number.

Prompt 7. CEO Weekly Financial Email

What it does. Compresses the weekly close into a five-bullet Friday email a busy CEO will actually read.

Draft a 5-bullet Friday close email from CFO to CEO. Tone: direct, no throat-clearing.

Inputs to synthesize:
- Variance summary: [PASTE PROMPT 1 OUTPUT]
- Cash forecast: [PASTE PROMPT 4 OUTPUT]
- Anomaly flags: [PASTE PROMPT 3 OUTPUT IF ANY]

Bullet order:
1. Revenue vs plan (one sentence with $ and %).
2. Gross margin movement (one sentence with driver).
3. Cash position and the one tight week ahead.
4. The single biggest anomaly worth the CEO knowing.
5. What I am doing about item 4 by end of next week.

Subject: "Weekly close, [DATE]"
Under 120 words. No sign-off.

Which LLM. GPT-5.

Common failure. Bullet 5 lands as “we will monitor the situation.” Reject and rerun with the constraint “must name a specific action, owner, and deadline.”

Scenario Modeling (Prompts 8 and 9)

Prompt 8. Three-Case Revenue Sensitivity

What it does. Turns a single-point revenue forecast into base, upside, and downside cases with named assumptions for each.

You are an FP&A analyst building a three-case revenue view for [PERIOD].

Base case revenue: $[X], built on:
- Volume: [UNITS]
- Price: $[PRICE]
- Mix: [DESCRIBE]
- Churn assumption: [%]
- New logo close rate: [%]

Produce upside and downside cases by moving 2 to 3 of these levers, not all of them.
- Downside: which 2 levers get worse and by how much? Give a plausible narrative.
- Upside: which 2 levers get better and by how much? Give a plausible narrative.

Output as a 3-column table (Downside, Base, Upside) with the 5 input rows above, then a Revenue total row.

Do not move any lever by more than 20% from base without a named external cause.

Which LLM. Claude Sonnet.

Common failure. Skipping the “no lever moves more than 20% without a named external cause” rule. Otherwise the downside case reads like a horror novel.

Prompt 9. Pricing Change Impact Model

What it does. Estimates revenue, gross margin, and cash impact of a proposed price change under different elasticity assumptions.

You are a CFO evaluating a pricing change.

Current state:
- Product/SKU: [NAME]
- Current price: $[P0]
- Current volume: [Q0] per period
- Current variable cost per unit: $[VC]
- Current gross margin per unit: $[P0 - VC]

Proposed change: [+/- X%] price move

Model 3 elasticity assumptions:
- Low elasticity (-0.3): what happens to volume, revenue, GP?
- Medium (-0.8): same 3 metrics.
- High (-1.5): same 3 metrics.

Output as a table with columns for each elasticity case and rows for: New Price, New Volume, Revenue, Gross Profit, GP % change vs today.

End with one line: at which elasticity does GP go negative vs today?

Which LLM. Either, but use one with a code interpreter/analysis tool so the arithmetic is calculated, not estimated.

Common failure. The model uses simple linear extrapolation instead of the elasticity formula. Spot-check by asking it to show the calculation for one case.

Interview Prep (Prompt 10)

Prompt 10. Finance Hire Interview Questions Grounded in Real Financials

What it does. Generates a set of interview questions for an FP&A hire, grounded in your actual financials so the candidate cannot pattern-match to generic answers.

You are helping a CFO interview an FP&A candidate for a [SENIORITY] role.

Company context: [SECTOR, REVENUE, ORG SIZE]
Real (redacted) monthly P&L: [PASTE, WITH DOLLARS SCALED 10x OR 0.1x FROM ACTUAL]

Generate 8 interview questions across 4 categories:
1. Reading the numbers (2 questions on what the P&L is showing)
2. Root cause hypotheses (2 on what could drive a specific variance)
3. Modeling (2 asking how they would forecast a specific line)
4. Judgment (2 on what they would recommend to the CEO)

Rules for each question:
- Reference a specific line item, region, or product from the input.
- Do not include the answer. But include a one-line grading rubric: what a strong answer covers.
- No "tell me about a time when" questions. Only questions about these numbers.

Which LLM. Claude Sonnet, for the grounding.

Common failure. Forgetting to scale the numbers before pasting. Never interview against un-redacted real financials.

How to Use This Library

Three practical notes on getting compounding value from the ten prompts above.

Put them in a Claude Project or a ChatGPT Custom GPT. Both let you attach a system prompt and reference files that persist across every conversation. Load your chart of accounts, business context paragraph, and materiality thresholds once, then run the prompts against that context every week. This is the difference between a prompt working the first time and it working every week.

Version the prompts. Save them in a shared “cfo-prompts.md” file with a version number and a date. When a prompt fails, edit and increment the version. Six months in, you have a private prompt library that gets sharper each quarter. This is the CFO equivalent of what engineers call prompt engineering.

Redact inputs. The privacy problem is real. If you are running these on a consumer tier, scale every dollar by a constant (multiply by 0.37 or similar) before pasting. If you are on Claude for Work, ChatGPT Enterprise, or M365 Copilot with commercial data protection, the vendor commits inputs are not used for training and you can paste unredacted. Do not confuse the two tiers.

Frequently Asked Questions

Can I run these prompts in one long chat, or should each be its own conversation?

Prompts 1, 3, and 4 should be separate conversations because you want a clean context per input. Prompt 7 (CEO email) works best in the same chat as prompts 1, 3, and 4 because it needs their outputs. Everything else, either way. Rule of thumb: separate conversations when inputs are unrelated, same conversation when the next prompt consumes the last output.

How much do I need to customize before these work?

Replace the bracketed placeholders, and adjust the business context sentence to reflect your sector, size, and gross margin profile. That is about 10 minutes of setup per prompt. Do not rewrite the “rules” section on your first run. The rules exist to prevent the specific failure modes I called out; you can iterate on them once you have seen how the model responds.

What if my company runs on Gemini or M365 Copilot instead of Claude and ChatGPT?

The prompts still work. Gemini needs slightly more explicit format constraints (“output only the table, no preamble, no explanation”) because its default is chattier. M365 Copilot works best when you attach the source Excel file as context and ask it to reference the file rather than pasting the table into the chat. The head-to-head comparison covers the per-tool differences.

Do these prompts replace an FP&A analyst?

No. They replace the first draft an analyst would spend two hours writing. The analyst then reviews, edits, and adds the context only they know (the sales team lost a big deal Tuesday, IT rebuilt a server so hosting costs spiked). Human plus prompt beats human alone and beats prompt alone; that is the story of AI in finance in 2026.

How do I know when a prompt is “good enough” to standardize?

Run it three weeks in a row on real data. If the output is usable with fewer than 5 minutes of editing all three times, standardize it. If you rewrite more than 30% of the output any of the three weeks, the prompt is not tight enough; identify what you keep rewriting and add a rule that prevents it.

Sources

Written by The Pragmatic CFO. 15+ years running P&Ls and AI-native finance experiments across portfolio companies.