I work in finance — the kind of work where a wrong number in cell F42 becomes a problem in a board meeting. For a long time I treated AI as a writing assistant and nothing more. Then I started writing structured prompts for the analytical work itself: explaining variances, sanity-checking models, drafting investor commentary, and turning a mess of department inputs into a coherent narrative.
These prompts don't replace the spreadsheet. They replace the hour I used to spend staring at a variance trying to find the words, or the hour I spent rewriting a memo because the first draft buried the only number the CFO cared about. Here are ten I use regularly.
How these prompts are structured
Each one has four parts:
- Role — a specific persona with expertise and a point of view.
- Context — the actual numbers, draft, or scenario (paste it in).
- Constraints — hard rules the output must obey.
- Output — the exact format I want back.
The constraints are where the value is. An unconstrained "explain this variance" prompt produces a paragraph of hedging. A constrained one produces a decision-ready answer.
1. The variance explanation that leads with the number
Role: You are an FP&A manager writing a variance commentary for a CFO who reads 40 of these a month. You lead with the number, never with the narrative.
Context: Actual was $[X]. Budget was $[Y]. The variance is $[X-Y]. Here is what drove it: [LIST THE DRIVERS — e.g., "software revenue came in $40k above plan due to 2 enterprise deals closing early; consulting revenue was $15k below due to a project slipping to next month"].
Constraints:
- First sentence must state the variance direction and dollar amount. Example shape: "Revenue came in $25k above plan (12% favorable)."
- Then list the drivers in descending order of dollar impact.
- Do not use the word "primarily" or "mainly" without a number attached.
- Do not open with "Overall," "In general," or "For the period."
- If a driver is unfavorable, say so plainly — no "headwinds," no "softness."
Output: 4-6 sentences. No paragraphs longer than 3 sentences.
This is the prompt that fixed my commentaries. The constraint "first sentence states the number" eliminates the 200-word throat-clearing that makes variance reports unreadable.
2. The model sanity check
Role: You are a financial modeler auditing a colleague's model for the five most common errors before it goes to the investment committee.
Context: Here are the key outputs of the model: [PASTE REVENUE PROJECTIONS, MARGINS, CASH BALANCE FOR YEARS 1-5]. Here are the key assumptions: [PASTE GROWTH RATE, PRICING, CAC, CHURN, etc.].
Constraints:
- Flag any year where revenue growth accelerates without a stated driver in the assumptions.
- Flag any year where gross margin moves more than 5 points without explanation.
- Flag if cash balance hits zero in any year (funding gap).
- Flag if CAC payback exceeds 24 months.
- Flag if the terminal growth rate exceeds the long-term GDP growth assumption.
- For each flag, state the specific cell/line and the specific concern in one sentence.
Output: A numbered list of flags. If no flags, say "No flags" — do not invent concerns.
The "do not invent concerns" constraint is critical. AI will manufacture problems to seem thorough. You want it to only surface real ones.
3. The investor memo draft (from bullet points)
Role: You are a CFO's chief of staff drafting the monthly investor update. You know investors read the first paragraph and skim the rest, so the first paragraph must contain the three things they actually care about.
Context: Here are this month's raw notes in bullet form: [PASTE MESSY BULLETS — revenue, hires, product, asks, lowlights].
Constraints:
- Paragraph 1 (the "headline paragraph"): revenue vs plan in one sentence, cash runway in months in one sentence, the single most important thing that happened this month in one sentence.
- Then four labeled sections: Highlights, Lowlights (mandatory — do not omit), Asks (specific, not "intros welcome"), Hiring.
- Each section: max 3 bullets, each bullet a complete sentence with a specific number or name.
- Do not use the words "excited," "thrilled," "pleased to announce," or "milestone."
- The Asks section must contain at least one specific request (a name to meet, a role to fill, a customer intro), not "we'd love introductions."
Output: The memo, ready to edit. No placeholder text.
Forcing the lowlights section is the constraint that builds trust. Investors discount updates that only report good news.
4. The board-deck commentary for a single slide
Role: You are a strategy consultant writing the speaker notes for a single board slide. The CEO will read these notes verbatim if nervous, so they must sound like a person, not a slide.
Context: The slide title is: [TITLE]. The slide shows: [DESCRIBE THE CHART — e.g., "a bar chart of ACV by customer segment for the last 4 quarters"]. The one point the CEO wants the board to take away is: [THE TAKEAWAY].
Constraints:
- Write 90-120 words of speaker notes.
- Open with the takeaway, not the data description ("What this shows is that mid-market ACV has stabilized" — not "This chart displays...").
- Reference one specific number from the chart.
- End with the implication or the ask, not with a description of the chart.
- No bullet points in the notes — flowing prose.
Output: The speaker notes.
5. The "what would kill this deal" red-team prompt
Role: You are a skeptical due-diligence analyst. Your job is to find the three reasons this investment would lose money, even if you're bullish on the business.
Context: Here is the investment thesis and the key metrics: [PASTE].
Constraints:
- Identify the three highest-velocity risks — the things that would go wrong fastest if they go wrong.
- For each risk, name the specific metric or assumption that would prove the risk is materializing (a "canary" metric).
- Rank by speed of onset, not by magnitude.
- Do not include generic risks like "competition" or "macro" unless tied to a specific assumption in the thesis.
- Do not hedge — state each risk as a conditional that could be true.
Output: Three risks, each with: the risk, the canary metric, and the onset speed.
The "specific canary metric" constraint forces specificity. "Competition" is not a risk; "if CAC rises 30% in two quarters while organic search stays flat, we've lost positioning" is.
6. The scenario commentary (bear / base / bull)
Role: You are a CFO explaining three scenarios to a board that wants to know which triggers would move you from one to another.
Context: Bear case: [SUMMARY + KEY NUMBER]. Base case: [SUMMARY + KEY NUMBER]. Bull case: [SUMMARY + KEY NUMBER].
Constraints:
- Write a 3-paragraph commentary.
- Paragraph 1: state the spread between bear and bull in one number (e.g., "$8M revenue spread").
- Paragraph 2: name the two variables that explain 80% of the spread.
- Paragraph 3: the specific trigger that would move the company from base to bear, and the specific trigger that would move it from base to bull — each tied to a measurable threshold, not a vibe.
- Do not say "we remain cautiously optimistic."
Output: Three paragraphs, labeled.
7. The KPI dashboard narrative
Role: You are a RevOps analyst writing the Monday-morning Slack summary of last week's metrics. The leadership team reads this on their phones before 9am.
Context: Here are last week's KPIs vs target: [PASTE — e.g., "MRR $420k vs target $415k; new logos 12 vs target 15; churn 1.2% vs target 1.5%; pipeline $1.9M vs target $2.2M"].
Constraints:
- First line: one word — "Green," "Yellow," or "Red" — based on whether the majority of KPIs hit target.
- Then exactly 3 bullets: the best metric (with the delta), the worst metric (with the delta), and the one metric most likely to move next week.
- Total length under 100 words.
- No commentary longer than one sentence per bullet.
Output: Slack-ready summary.
This replaced a 400-word Monday email nobody read. The "one word first" constraint forces a verdict instead of a hedge.
8. The financial glossary entry (for non-finance readers)
Role: You are a finance educator writing for smart non-finance people — engineers, designers, founders who need to understand a term to do their jobs, not to pass the CFA.
Context: Define this term for an internal wiki: [TERM].
Constraints:
- Definition in one sentence, plain English, no jargon in the definition itself.
- Then a 2-sentence example using a concrete small-business scenario (not a public company).
- Then: "What it is NOT" — one sentence clarifying the most common confusion.
- Max 80 words total.
Output: The glossary entry.
The "what it is NOT" constraint is the one that makes these useful. Most confusion in finance is between adjacent terms (EBITDA vs cash flow, gross margin vs contribution margin) and the disambiguation is what actually teaches.
9. The quarterly business review (QBR) narrative
Role: You are a consulting manager drafting the narrative arc of a QBR. You know the audience (a client's leadership team) will remember the story, not the slides.
Context: Here are the quarter's results and themes: [PASTE — wins, misses, strategic shifts, next-quarter focus].
Constraints:
- Write a 4-section narrative.
- Section 1 ("Where we said we'd be"): restate last quarter's commitments and mark each met / missed / partial.
- Section 2 ("Where we are"): the actual results, leading with the miss if there was one (do not bury it).
- Section 3 ("What changed"): the 1-2 strategic shifts made this quarter and why.
- Section 4 ("Where we're going"): next quarter's 3 commitments, each measurable.
- No section longer than 150 words.
Output: Four labeled sections.
Leading with the miss in section 2 is the constraint that makes QBRs credible. Burying a miss in slide 14 is how trust erodes.
10. The "translate this for the CEO" prompt
Role: You are a chief of staff translating a technical finance memo into language a non-financial CEO can act on in the next 10 minutes.
Context: Here is the memo: [PASTE].
Constraints:
- Produce a 5-bullet summary.
- Bullet 1: the decision being asked for, in one sentence.
- Bullet 2: the cost or risk of the decision, in one number.
- Bullet 3: the cost or risk of NOT deciding, in one number.
- Bullets 4-5: the two most important caveats.
- Remove every piece of jargon. If "EBITDA" or "net working capital" appears, replace with the plain-English equivalent on first use.
Output: Five bullets. No preamble.
This prompt exists because too many memos are written for the author, not the reader. The constraint "the decision in one sentence" forces the memo author to know what they're actually asking for.
A note on using AI with financial data
Be deliberate about what you paste. I never paste customer names, deal-specific pricing, or anything contractually confidential into a third-party model. I paste structures, formats, and numbers with labels stripped or generalized. The prompts above are designed to work with the shape of the analysis — they don't need the actual sensitive figures to be useful. A variance explanation prompt works whether the variance is $25k or $2.5M; the structure is the same.
If your organization has an approved internal AI environment, use that for anything sensitive. If not, sanitize first.
How I test a finance prompt
Three runs on the same input. I'm looking for structural consistency (the output always leads with the number, always lists drivers by magnitude) but acceptable variation in phrasing. If the output changes its structure between runs, the prompt's constraints are too loose for work where consistency matters. Tighten until the structure is reliable, then trust it.
The goal isn't to remove judgment — it's to remove the blank-page tax. Every one of these prompts replaces 20-60 minutes of "where do I start" with a draft I can edit in five.
If these were useful, I maintain a larger library of tested prompts for the full range of operational work — product management, engineering management, marketing, customer support, data analysis, recruiting, sales, and more. The Developer Product-Launch Prompt Pack is the $9 entry point, and the full set is on my Gumroad storefront.
Every prompt in the packs uses the same Role → Context → Constraints → Output structure and the same emphasis on hard constraints over verbosity.
Top comments (0)