Analytical Prompt Library, Week 6

AI is most useful when you drive it, and driving it well is a skill. These are categories of prompt that move an analysis from raw data to a finding you can stand behind, in order: get oriented, name the shape, read a chart someone hands you (or that you are about to present), hunt the anomaly, stress-test the finding, then communicate the story. Each prompt has a copy button. The bracketed parts, like [column], are yours to fill in. Read the "why it works" line under each category, because the reasoning transfers to prompts this page never wrote.

The rule that ties them together: never paste a number into a slide that the AI produced and you did not check. AI is a thinking partner for the analysis, not a source of truth for the figures. You own every number you present.

1 Explore 2 Characterize 3 Read a chart 4 Hunt the anomaly 5 Stress-test 6 Communicate Why prompts work

1Explore: get oriented before you analyze

Goal: understand what is in the dataset before you commit to an analysis, so you do not discover a broken column three slides in.

Orient on a new dataset

Here are the columns in my dataset: [paste the column list]. Before I analyze anything, what should I check first? List the data-quality issues to look for (types, ranges, missing values, obvious anomalies) and tell me which columns are likely to matter most for [the question I am trying to answer].

Suggest the first cuts

I have a table of [what the rows are, e.g. invoices] with these columns: [list]. What are the three or four summaries I should run first to get oriented? For each, say what I would learn and what would count as a surprise worth investigating.

Why it works: you surface type, range, missing-value, and anomaly issues up front instead of hitting them late, after you have already built on a column that turns out to be half empty. Giving the AI the actual column list, rather than asking "how do I explore data," is what makes the answer specific to your problem.

2Characterize the distribution: name the shape, pick the right summary

Goal: decide whether to report the mean or the median, and which chart shows the shape, before you summarize.

Mean or median, and why

For the column [column], here is a summary: mean [value], median [value], min [value], max [value], and the rough shape is [what you see in the histogram]. Is this column symmetric or skewed? Should I report the mean or the median, and why? Which chart would best show its shape to a non-technical audience?

Name the distribution

My column [column] has mean [value] and median [value], and the histogram shows [describe it: one hump, a long right tail, two humps, etc.]. Which named shape is this closest to (normal, right-skewed, long-tailed, bimodal, or Pareto)? What does that shape imply for how I should summarize and present it?

Why it works: it forces the center-and-shape decision this whole week is about, instead of letting you default to the mean out of habit. By handing the AI the mean, the median, and the rough shape, you give it exactly what it needs to tell you the mean is misleading here, the gap between the two numbers is the tell.

3Read a chart: paste the picture and ask the model what it shows

Goal: hand a multimodal model a screenshot of a chart and have it read the chart back to you, then flag anything misleading, so a second set of eyes catches what you missed.

Read this chart for me

[Paste a screenshot of the chart.] Read this chart for me. In plain language, what does it show, what is on each axis, and what is the one main takeaway? Then tell me what it does NOT let me conclude.

What is misleading here

[Paste a screenshot of the chart.] Look at this chart critically and flag anything misleading. Check the y-axis: does it start at zero, or is it truncated to exaggerate a difference? Is the window cherry-picked to a flattering date range? Is a trend implied that the data does not support? List what an honest version would change.

Why it works: a multimodal model can see the chart, so it can sanity-check your read and catch a misleading axis you skimmed past. A truncated y-axis or a cherry-picked window makes a small change look dramatic, and those are exactly the tricks a second reader spots fastest. Use it both ways: to read a chart you are unsure of, and to pressure-test one you are about to present, so you find the truncated axis before your audience does. Still check the read against the underlying numbers, the model can misread a label.

4Hunt the anomaly: find what does not fit

Goal: surface the rows that break the pattern, because in audit and fraud work the outlier is often the finding.

Find the rows that break the pattern

Here is a sample of my data: [paste a small sample or describe the columns and ranges]. Find the rows that break the pattern. For each one, tell me why it stands out and whether it reads as a likely error (a typo, a unit mix-up, a duplicate) or a real signal worth investigating.

Error or signal

I found these unusual values in [column]: [list them]. For each, give me the most likely innocent explanation and the most likely concerning explanation. What single follow-up check would tell the two apart?

Why it works: the outlier is frequently the whole point of the analysis, the duplicated payment, the journal entry posted at 2 a.m., the customer whose balance does not add up. Asking the AI to separate "error" from "signal," and to name the check that distinguishes them, keeps you from either ignoring a real finding or chasing a typo.

5Stress-test the finding: make the AI argue against you

Goal: kill the rationalization before it reaches a slide. The point is not agreement, it is the strongest counter-argument.

Argue against my claim

I am about to claim: [your finding, in one sentence]. Argue against me. What would make this wrong? What is the strongest counter-argument? What did I likely fail to control for, and what alternative explanation fits the same data?

Pre-mortem the conclusion

Assume the conclusion [your finding] turns out to be wrong when someone checks it next quarter. Walk backward: what was the most likely reason it was wrong? Which assumption was the weakest link, and what evidence would I need to shore it up before I present it?

Why it works: an AI that is trying to agree with you will rationalize a shaky finding right onto your slide. Explicitly telling it to argue the other side turns it into the skeptical reviewer you want before the meeting, not after. This is the Week 5 analytic-argument habit applied to your own conclusion.

6Communicate the story: turn the finding into a narrative

Goal: the present-information payoff. Lead with the conclusion, one message per slide, the memo a decision-maker can act on.

Three-sentence executive summary

Here is my finding and the evidence behind it: [finding, plus the key numbers]. Turn this into a three-sentence executive summary that leads with the conclusion, not the methodology. Sentence one is the bottom line, sentence two is the evidence, sentence three is the recommended action.

The one slide a CFO needs

From this finding: [finding and numbers], draft the single slide a CFO needs. Give me a headline that states the conclusion, the one chart that proves it (name the chart type and what is on each axis), and the one number to put in large type. Keep it to one message.

Translate for the audience

I need to present [finding] to [audience: a CFO / an auditor / a non-financial manager]. Rewrite it for that audience: what do they care about, what jargon should I drop, and what is the one action I should ask them to take?

Why it works: leading with the conclusion (bottom line up front) is what separates an executive summary from a lab report. Asking for one message per slide and one number in large type forces the discipline that makes a deck land. Naming the audience changes what matters, the same finding reads differently to a CFO and an auditor.

Weak prompt

Write a summary of my receivables analysis.

Vague. No audience, no finding, no shape to the output. You will get a generic paragraph that buries the point and that you will have to rewrite anyway.

Strong prompt

My finding: overdue dollars are concentrated in the commercial segment, where 12 accounts hold 80 percent of the past-due balance, and the balances are right-skewed so the median overdue account is small. Write a three-sentence executive summary for a CFO that leads with the conclusion, then the evidence, then one recommended action.

Specific finding, the shape named, the audience named, the output structure given. You get a summary you can almost paste, after you check the numbers.

7Why prompts work: the patterns underneath

Every prompt on this page is built from the same handful of moves. Learn the moves and you can write prompts for analyses this page never imagined.

Be specific. "Analyze my data" gets you nothing. The column list, the actual numbers, the question you are answering, that is what makes the response yours instead of generic.
Give the data shape. Hand the AI the mean, the median, the min, the max, and what the histogram looks like. It cannot see your screen, so the more shape you give it, the better it reads the distribution.
Ask for reasoning, not just an answer. "Why" and "what would make this wrong" pull out the thinking you can check, instead of a number you have to trust on faith.
Make it argue against you. An AI that agrees is not testing your finding. Tell it to take the other side and it becomes a reviewer instead of a cheerleader.
Never paste a number you have not checked. The AI can shape the story and sharpen the prose. The figures on your slide are yours to verify. Use it to think, not to source the facts.

Carry this off the page: these five categories are a workflow, not a menu. Explore to get oriented, characterize to pick the right summary, hunt the anomaly, stress-test before you commit, then communicate. The prompts are starting points; the five principles above are what let you write your own for any dataset, in any tool, long after this course.

ACCTG 5150, Summer 2026, Week 6. Pair this with the Distributions Field Guide so you can name the shape the prompts ask the AI to characterize. Brackets like [column] are yours to fill in.