Level 7 · 2 to 3 hours

Analyze a report or simple dataset

You will analyze a report or a simple dataset you are allowed to share, or a small table you type and clearly label as practice data. Ask Gemini to separate what the figures show from what they might mean and from what they cannot prove. Check at least one calculation yourself, and keep every recommendation labeled as a hypothesis until a source supports it. Do not upload confidential files, and do not ask the model to see numbers you never provided. Screenshots and tables still need you to compare the reading with the original. The result is a short decision memo: observations, limits, and the next question a human should answer before anyone acts.

Outcomes

A source note that says what you analyzed, what you left out, and why you were allowed to share it.
A memo that separates observations, hypotheses, and things the data cannot prove.
One calculation you checked by hand, plus a next question a person still has to answer.

Steps

1. Choose a source you may share
Use a report or table you are allowed to share, or type a small practice table and label it practice data so nobody confuses it with a real company result. Remove names, account numbers, and anything confidential before it touches the chat. Write one sentence on what decision the memo might inform. If you cannot remove the sensitive parts, pick a different source.
2. Ask for a reading, not a strategy
If you use a screenshot or a chart image, ask first for the labels and figures that are actually visible. If you paste a table, ask the model to restate the columns before interpreting them. Compare that reading with the original yourself. A mismatch means you stop and correct the figures. Do not let a recommendation run ahead of a reading you have checked.
3. Separate observation from story
Request three labeled parts: what the figures show, what they might mean, and what they cannot prove. Trends, causes, and advice belong in the second or third part unless the source itself states them. Ban invented context, such as a campaign or a competitor you did not mention. A fluent explanation of why a number moved is still a hypothesis until something else supports it.
4. Check one calculation yourself
Pick one total, difference, or simple comparison and recompute it without the model. If you and the draft disagree, trust the arithmetic you can show and correct the memo. Note the check in one line so a reader knows you did it. Skipping this because the paragraph sounded sure is the failure this project is designed to catch.
5. Write the decision memo
Close with a short memo a busy person could use: the source, three observations, the limits, and one question a human must answer before acting. Recommendations stay marked as hypotheses. Do not add a confident plan that the table cannot carry. Save the original figures beside the memo so you can audit a line later without reopening a long chat.

Starter prompts

Read the source first

I will paste a small table or upload a chart I am allowed to share. If I say it is practice data, treat it as practice and do not invent a real company around it. First, restate only the labels and figures you can see. Do not interpret yet. If a number is unclear, say unclear rather than guessing. After the reading, stop and wait for me to confirm that it matches the original. Do not recommend an action, do not add causes, and do not import numbers I did not provide. Private names and account details should not appear; if you notice them, tell me to remove them.

Decision memo

Using only the figures I confirmed, write a short decision memo with these headings: Source, Observations, Hypotheses, What this cannot prove, and Next human question. Observations must mention only relationships that are visible in the confirmed figures. Put causes, advice, and forecasts under Hypotheses. Do not invent context, customers, or extra months of data. Leave the recommendation unstated if the table cannot support one. End by naming one calculation I should recompute myself, and do not claim the analysis is audited until I say I checked it.

Checklist

The source is either shareable or clearly labeled practice data, with secrets removed.
I compared the model's reading of the labels and figures with the original.
Observations, hypotheses, and limits are in separate parts of the memo.
I recomputed one calculation myself and noted the check.
No cause or recommendation is written as if the table proved it.
The memo ends with a question a human still has to answer.

Rubric

A reader can tell what was in the source and what was left out.
The reading matches the original labels and figures.
Stories about why the numbers moved are labeled as hypotheses.
At least one calculation was checked by you, and the next decision is still a human question.