Most advice about getting AI to read data is abstract: "be specific," "verify the output." True, but not actionable when you have a messy table in front of you and twenty minutes to produce an answer. This article is the opposite. It is a sequence of concrete steps you can follow start to finish today, in order, with nothing skipped.
The process works for both spreadsheets pasted as text and charts provided as images, with a few branches noted where the two diverge. It assumes you want an answer you can defend, not just one that sounds plausible. That means a little setup and a verification pass at the end—both built into the steps below.
Follow it in order the first few times. Once the moves are habitual, you will compress several steps into a single well-formed prompt. But the order encodes the reasoning, so learn it sequentially before you start shortcutting.
Step 1: Prepare the Data Before the Model Sees It
The work you do before prompting determines most of the outcome. Skipping this step is the most common reason answers go wrong.
For text tables
- Paste the table and check that columns stayed aligned.
- Confirm every column has a clear header, including units.
- Remove decorative rows, merged cells, or totals that could be mistaken for data.
For chart images
- Make sure axis labels and the legend are legible in the image.
- Note the scale yourself—linear or logarithmic—because it changes every reading.
A clean starting point prevents the structural misreads detailed in the common mistakes piece.
Step 2: State What the Data Represents
Before asking your question, tell the model what it is looking at. One or two sentences of context dramatically improves interpretation.
What to include
Say what the rows are, what the columns measure, the units, and the time period if relevant. "This table shows monthly new signups by plan tier for 2024; counts are whole numbers." This single sentence eliminates an entire class of guesses, because the model no longer has to infer what the numbers mean.
This framing step is the same instinct behind the structured stages in the data interpretation framework.
Step 3: Ask One Specific, Checkable Question
Resist the urge to ask everything at once. Start with a single question that has a verifiable answer.
How to phrase it
Aim for a question whose answer is a number or a named cell. "Which plan tier had the most signups in March, and what was the count?" is checkable. "Summarize the signup trends" is not. You can ask the broad question later, once you have confirmed the model is reading the table correctly.
Step 4: Request the Reasoning Alongside the Answer
Ask the model to show which cells it used and any calculation it performed. This turns the answer into something you can audit in seconds.
What good reasoning looks like
For a computed value, you want to see the formula and the inputs: "March, Pro tier: 1,840 signups, read from row 3, column 'Pro.'" If the model gives a number without showing where it came from, ask again. The reasoning is what lets you catch a wrong-cell error before it spreads.
Step 5: Verify Before You Trust
Now do the check. This is the step people skip and later regret.
The verification pass
- Find the model's cited cells in the source and confirm the values.
- Recompute any headline number yourself.
- For image data, confirm the model read the axis scale correctly.
- Ask the model to flag anything it estimated or was unsure about.
If everything holds, you have an answer you can stand behind. If not, you have caught the error privately instead of in front of a client. Teams that do this at volume formalize it, as shown in the data prompting case study.
Step 6: Expand to Harder Questions
Only after the model has proven it reads the table correctly should you ask for trends, comparisons, or summaries.
Going deeper safely
- For trends, ask the model to quantify the change between periods, not just describe direction.
- For comparisons, have it state both values and the difference.
- For summaries, ask it to ground each claim in specific cells.
This sequencing—confirm accuracy first, then interpret—keeps the harder questions anchored to verified ground, the same principle emphasized in the best practices guide.
Step 7: Handle Calculations Explicitly
Many real questions require arithmetic—growth rates, shares of a total, differences between periods. This is where confident errors are most common, so the step deserves its own discipline.
Make the math visible
For any computed value, ask the model to state the formula and plug in the actual numbers before giving the result. A growth-rate answer should read "Q4 minus Q3, divided by Q3: (1,840 minus 1,600) / 1,600 = 15 percent," not just "15 percent." With the inputs shown, a wrong cell or a slipped decimal is obvious at a glance.
Recompute what matters
For any number that will appear in a report or drive a decision, do the calculation yourself once. This takes seconds for a single metric and is the cheapest insurance against the kind of error that survives because it looks tidy. Models are competent at arithmetic but not flawless, and the cost of a wrong headline number is high.
Step 8: Document the Prompt That Worked
Once you have a prompt that produces a reliable, verified answer, save it. The next similar table will go far faster.
Why this compounds
A working prompt for "monthly metrics by category" can be reused, lightly adapted, the next time you face a similar table. Over a few weeks you accumulate a small library of prompts that already encode the context, the checkable question, and the request for shown work. This is how individual effort turns into a repeatable team capability, much like the standardized workflow in the data prompting case study.
What to capture
- The context sentence describing the data.
- The question phrasing that produced a checkable answer.
- The instruction to show cells and calculations.
- Any data-specific rules you had to state, like multiple-select or unit scale.
Adapting the Sequence Under Time Pressure
The full eight steps are for when accuracy matters most. Under real deadlines, you compress without abandoning the load-bearing parts.
The compressed version
When you are short on time, fold preparation, context, and the question into a single prompt, and keep only two checks: recompute the headline number and sanity-check its size. This trims the process to a couple of minutes while preserving the steps that catch the costliest errors. The full sequence remains the standard for anything client-facing, as the common mistakes piece makes clear.
Frequently Asked Questions
How long does this process actually take?
The first few times, maybe ten minutes for a moderate table. Once the steps become habit, you fold preparation, context, and the question into one prompt and run the verification pass in two or three minutes. The setup time pays for itself the first time it catches an error.
Can I combine the steps into a single prompt?
Yes, once you are practiced. A mature prompt states what the data represents, asks a specific question, and requests cited reasoning all at once. The reason to learn the steps separately first is that the order encodes why each move matters; combining them too early hides the logic.
What if the model refuses to show its reasoning?
Ask directly: "Which rows and columns did you use, and what calculation did you perform?" If it still gives only a bare answer, treat that answer as unverified and recompute it yourself. A model that cannot show its work has not earned your trust on that number.
Does this work for very large tables?
For large tables, the same steps apply but verification matters more, since you cannot eyeball the whole thing. Spot-check several cited cells rather than one, and recompute any aggregate. Consider breaking the question into parts so each answer stays checkable.
Should I do anything different for charts versus tables?
Two things. First, confirm the chart's scale yourself before trusting any reading. Second, expect approximate values from images and ask for ranges rather than precise figures. Otherwise the sequence is identical: prepare, contextualize, ask, request reasoning, verify.
Key Takeaways
- Most of the outcome is determined before you prompt—clean and prepare the data first.
- State what the data represents in one sentence to eliminate a whole class of guesses.
- Ask one specific, checkable question before any open-ended summary.
- Always request the cited cells and calculation so the answer can be audited in seconds.
- Confirm the model reads the table correctly, then expand to trends and comparisons anchored to verified numbers.