Give it the data, not a picture of the data

What the move is

The single highest-value substitution in analysis work. A CSV removes an entire measured failure mode that no prompt fixes.

Why it works

A benchmark built around realistic web work found visual reasoning over rendered content — not clicking — to be the dominant failure mode, with near-zero performance on reading precise values off charts and tables.

Step by step

1. Before uploading an image of data, ask where the underlying file or endpoint is.
2. Send the CSV, the table, or the query result.
3. Where only an image exists, ask for axis labels and gridline values first, then the point, and check them.
4. Prefer 'write code that parses this file' over 'read this screenshot'.
5. Treat any figure that arrived through a screenshot as unverified until you have seen it in the source.

Paste this

Attached is the file the chart was built from. Parse it with code and show me the parsed rows before any analysis, then answer <the question> from those values. If a number is not in the file, say so.

Make it yours

If the data is behind a dashboard with no export, screenshot it and transcribe the numbers yourself once. Transcribing ten numbers is faster than checking a hundred the model guessed.

Variations

• Ask it to write the parsing code and show you the parsed rows before any analysis.
• Give it the schema of the table alongside the file so column meanings are not inferred.

Works with

• Read a 600-page PDF including its charts
• Let the model write and run real code mid-answer