AI Misreading Your Table? Prompt for Source-Traced Numbers
Requires row and column coordinates, unit and period for every number, and marks anything unreadable as "uncertain". Essential for financial analysis.
You are going to read the document/table below. Your biggest risks are shifting rows and confusing units (thousands/millions/USD/EUR). We are going to prevent both. DOCUMENT: [DOCUMENT] MY QUESTION: [QUESTION] STEP 1 — DESCRIBE THE STRUCTURE How many pages/sheets? On which row does the header sit on each? Where is the unit stated (header, cell note, below the table)? What is the date range? Get me to approve this description. STEP 2 — SOURCE EVERY NUMBER Every number you use, in this format: | Value | Page/Sheet | Row | Column | Unit | Period | If you cannot pin down the row or column, DO NOT USE the number. List it as "uncertain" instead. STEP 3 — CONSISTENCY CHECK - Do totals match the sum of subtotals? - Does the same metric differ across pages? - Is there unit mixing (some cells in USD, some in thousands)? - Percentages or percentage points (15% vs 15 points)? If there are conflicts, name the exact cells. STEP 4 — ANSWER Answer only with numbers whose coordinates were verified in step 2. If you calculated, show the formula (e.g. "C4 - C5") so I can check it. STEP 5 — UNCERTAINTIES List separately what you could not answer, could not read or found contradictory. Do not fill gaps with guesses — a missing answer beats a wrong one.
On long PDFs the model concentrates on the early pages. Split the document into page ranges and run each separately.
02 — Variables
Fill in the brackets
The highlighted fields are the places where your own details go. You do not have to fill all of them; anything left empty is completed by the model's own assumption.
[DOCUMENT]
[QUESTION]
03 — How to use
How to use it, step by step
- 01
Never feed a large document in one go. Do it in 5-10 page chunks, one run per chunk.
- 02
Do not continue before approving the structure in step 1. If the model picks the header up from the wrong row, every number below is shifted.
- 03
For every numeric answer, ask "which cell?". If it cannot give coordinates, the number is invented or misread.
- 04
Always read the consistency report — real tables often contain genuine errors that are not the model's fault.
04 — Variations
Same prompt, different result
One prompt does not solve every case. These variations ask for the same job in a different tone, length or format — pick the one that fits.
Convert to spreadsheet
Extract the tables in the document as CSV. Rules: one block per table, header row first, add a unit column, leave empty cells empty (do not write 0), repeat merged cells on every row. If a table cannot be read, leave it blank and add "COULD NOT READ". Document: [DOCUMENT]
To turn the table into structured data.
Conflicts only
Do not summarise. Run a consistency audit only: totals, units, periods, percentages. Give me every conflict with cell references. If you find none, say "none found" — do not invent one. Document: [DOCUMENT]
For audit work.
05 — Tips
Notes that take the result up a level
- Scanned tables in PDFs do not become text; the model sees an empty page. Check for a text layer first.
- Unit mixing is the most common error: a "thousands" header is routinely ignored and the result comes out 1,000× too large.
- Percentages versus percentage points: going from 15% to 18% is 3 points but a 20% increase. Models confuse them; make the prompt ask separately.
06 — FAQ
Frequently asked about AI Auditing
Why does AI misread tables?
PDF tables are defined by visual position, not cell boundaries. The model takes text in reading order, can shift columns, does not repeat merged cells and loses the header across page breaks. The numbers look right and are attached to the wrong row.
What should I upload instead?
CSV or Excel where possible — cell structure survives and the error rate drops sharply. If you must use PDF, use a text-layer version rather than a scan. For table-heavy work, the reliable route is exporting to a spreadsheet first.
How do I analyse a table with a million rows?
Start with summary statistics: sum, mean, min/max, outliers for a given column. The model cannot read a million raw rows, but if it can run code, have it write and execute the analysis. The most reliable approach is to give the data to code, not to the model.
Next
You may also find these useful
Prompt to Catch Fake Citations: When the AI Invents Sources
Tests every source on four axes (does it exist, does it support the claim, is the metadata consistent, is it accessible) and surfaces the suspect ones first.
Intermediate ChatGPTClaudeAI Plagiarism Check: Rank Your Text by Risk Before Submission
Scans definition sentences, opening paragraphs and conclusions separately, then gives a source type and a rewrite direction for each risky sentence.
Advanced ChatGPTClaudePrompt to Make AI Text Sound Human (Remove the Tells)
Finds the clearest AI tells — rule-of-three balance, empty adjectives, cliché openers, connective overload — and rewrites around them without losing meaning.
Intermediate ChatGPTClaudePrompt to Remove Translationese From AI Output
Splits long sentences, turns passive into active, clears pronoun clutter and rebuilds the rhythm. For marketing and SEO copy.
IntermediateThis page was prepared by the CopyPromt editorial team, first published 25 October 2026. Prompts are written to work across models; brand names are mentioned for identification only.