Finance and Accounting Prompt Library

Prompts for finance work, and the verification rules that make them safe to use.

The rules

1. Never let a language model do arithmetic. It predicts numbers rather than calculating them. Every calculation goes through a tool that executes code, and you ask to see the working.

2. Reconcile to a known figure. Every output must tie back to something you already know is right — a bank balance, a control total, a prior period.

3. Never paste identifiable client financial data into a general tool. Anonymise, or use an approved tier with a processor agreement.

4. AI drafts; a qualified person signs. Nothing here substitutes for professional judgement, and in regulated work the accountability is yours regardless of what produced the draft.

Analysis and narrative

The commentary

Use code to calculate, not estimation. Show your working.

Here is the data: [PASTE]

1. Calculate the variances against [BUDGET / PRIOR PERIOD]
2. Identify the five largest by absolute value and by percentage
3. For each, state what the number is — not why. You do not know why.
4. List what I would need to check to explain each one

Do not speculate on causes.

That third instruction is the important one. AI will happily invent plausible business explanations for a variance, and a plausible wrong explanation in a board pack is worse than no explanation.

The narrative draft

Here are the figures and the explanations I have verified: [PASTE]
Audience: [BOARD / LENDER / MANAGEMENT]
What they need to decide: [DECISION]

Write the commentary.
- Lead with what it means for the decision
- Every number in context — a percentage needs its base
- State the one thing that most affects next period
- Flag anything I have marked as unexplained rather than smoothing over it

Use only the explanations I provided. Do not add interpretation.

Ratio and trend work

Use code. Calculate [RATIOS] for [PERIODS].
Show the formula used for each.
Identify any that moved more than [X]% and state the movement.
Flag any where the denominator changed materially — that often explains
the ratio without anything real having changed.

Reconciliation and cleaning

Finding the difference

Use code. Here are two datasets that should reconcile: [PASTE]

1. Total each and show the difference
2. Find items in one and not the other
3. Find items in both with different values
4. Find potential matches with formatting differences —
   dates, trailing spaces, case, reference formats

Present as a table. Do not adjust anything.

Cleaning, safely

Use code. Clean this data:
- Standardise dates to YYYY-MM-DD
- Strip whitespace and normalise case in [COLUMNS]
- Flag duplicates on [KEY] — list them, do not delete
- Leave blanks as blanks. Do NOT impute any value.

Show the code. List every change with a count.
Confirm the column totals before and after are identical.

That last line is the control. If a total changed during cleaning, something was destroyed.

Documents

Contract and agreement review

Here is an agreement: [PASTE]

Extract into a table: party, obligation, deadline, financial consequence,
termination condition, notice period.

Quote the clause reference for each row.
If something is ambiguous, mark it AMBIGUOUS rather than interpreting it.
This is a first pass for a qualified reviewer, not advice.

Policy questions

Here is our policy document: [PASTE]
Question: [QUESTION]

Answer using only this document, quoting the relevant clause.
If it does not address the question, say so — do not answer from
general accounting knowledge.

The last instruction matters enormously. Without it you get a generic answer about what is usually done, presented as your policy.

Forecasting

Use code. Here is [N] periods of actuals: [PASTE]

1. Calculate the trend and state the method
2. Identify seasonality if the data supports it
3. Project forward [N] periods
4. State every assumption explicitly
5. Give a range, not a point estimate
6. State what would invalidate this

Do not present a projection as a prediction.

Verification, every time

• Every number produced by executed code
• Working shown and checked on one line
• Totals reconcile to a known control figure
• Nothing imputed and presented as observed
• Percentages show their base
• No causal explanation that I did not verify myself
• A qualified person has reviewed anything going externally

Where AI does not belong

• Final figures in statutory accounts or filings
• Tax positions, without professional review
• Anything where the regulator expects a documented human judgement
• Client-identifiable data in an unapproved tool
• Audit evidence

The failure to watch for

Not the obviously wrong answer. The one that is plausible, well-formatted, confidently explained, and off by a factor nobody checks because the presentation is good.

The defence is unglamorous and reliable: reconcile one total, verify one line, state one uncertainty.