What this covers
You screenshot an error dialog, a tax form, a chart in a PDF, or a whiteboard from a meeting, and ask ChatGPT what it says. Half the time the answer is fine. The other half, it confidently misreads a number, omits a row, or fabricates a label that was never on the chart. The fix isn’t to stop using Vision — it’s to prompt it in a way that surfaces uncertainty instead of papering over it. This guide is for people who lean on Vision daily and want fewer silent OCR errors and chart hallucinations.
Who this is for
Engineers debugging from screenshots, analysts pulling numbers out of charts that exist only as images, support agents triaging error dialogs from customers, students explaining textbook diagrams, anyone who’d rather upload a picture than retype the contents. If you’re using Vision for “describe this artistically,” none of this applies — go play. If you’re using it for facts that have to be right, keep reading.
When to reach for it
- Transcribing a screenshot where retyping would take 10+ minutes (a tax form, a long error log).
- Explaining a chart, diagram, or UI flow to a colleague who’d benefit from a written summary.
- Triaging an error dialog to find the actionable line buried in stack-trace noise.
- Extracting handwritten notes from a whiteboard photo into a structured outline.
Before you start
- Crop the screenshot to the part you actually care about. Vision attention degrades on cluttered, full-screen captures.
- For small text (logs, code, fine print), zoom in before screenshotting — Vision’s OCR fails most often on text under roughly 12px.
- Decide upfront whether you need verbatim transcription or interpretation. The prompts are different and so are the failure modes.
- For anything where the numbers have to be exact (financial statements, dosing tables), plan to verify the model’s reading by spot-checking 3-5 values against the image yourself.
Step by step
-
Upload the image. Wait for it to actually attach — sending the prompt before the upload completes silently drops the image.
-
State what kind of read you want. “Transcribe” is different from “explain” is different from “extract”:
Transcribe the visible text in this screenshot verbatim. Preserve line breaks. Mark any character you can't read clearly as [?]. Do not infer text that isn't visible. -
For charts, ask for what you can verify (axes, labels, the shape of the trend) and explicitly do NOT ask for exact y-values it can’t measure:
Describe this chart: axis labels, title, legend, and the shape of each series. Do NOT estimate specific y-values from the visual — only report values that are explicitly labeled on the chart. -
For UI screenshots, ask the model to walk through what the user is being asked to do, step by step:
This is a screenshot of a settings page. List every action the user could take here, in the order they appear on screen. Don't infer behavior — only describe what's visible. -
For handwritten notes, expect ambiguity. Ask the model to surface its guesses:
Transcribe these handwritten notes. For any word you're uncertain about, give your best guess followed by [?] and 1-2 alternative readings. -
Re-upload at a higher resolution if the model says it can’t read a region — don’t make it guess.
A prompt that surfaces uncertainty instead of hiding it
Vision task: \{transcribe|explain|extract\} this image.
Constraints:
- Anything you can't read with confidence, mark with [?].
- Do not invent text, numbers, or labels not visible.
- If a chart value isn't explicitly labeled, do not
estimate it from the visual.
- If the image is too low-resolution for the task,
say so and stop instead of guessing.The “say so and stop” clause is the one that matters most. Without it, the model will produce a plausible-looking output instead of telling you the image was unreadable.
Quality check
- For verbatim transcription: spot-check 3-5 lines against the image. If any are wrong, re-prompt with “be more careful with line N” or upload a higher-res crop.
- For charts: check that the trend description matches what you see. If specific numbers came back, verify each one against an explicit label on the chart.
- For UI explanations: confirm every button or field mentioned is actually visible. Phantom buttons are a known failure mode.
- For handwriting: read the original yourself for any critical line. The model handles printing well, cursive less well, and abbreviations badly.
How to reuse this workflow
- Save a
vision-prompts.mdwith your three core variants (transcribe verbatim, explain chart conservatively, walk through UI). - For recurring image types — same dashboard, same form, same whiteboard layout — keep a per-type prompt that calls out the fields you always need.
- For high-stakes extraction (tax forms, contracts), build a verification checklist alongside the prompt — fields that must be re-read manually before the data is used.
Recommended workflow
Crop tight → upload → state read type (transcribe / explain / extract) → use uncertainty-surfacing prompt → spot-check key values against the image → re-upload higher-res if reads are weak.
Common mistakes
- Uploading a full-screen screenshot when you only care about one window. Vision gets distracted by the menu bar and tabs.
- Asking “what does this chart say?” without constraints. The model will read off the trend AND invent specific numbers that aren’t labeled.
- Trusting Vision on small text. Logs, code, and footer fine-print are where OCR errors silently slip in.
- Letting the model fill in fields that are partially cut off in the screenshot. It will produce plausible completions that don’t match the original.
- Using a low-resolution screenshot from a chat-app preview. Re-export from the source app at full resolution.
- Asking for both transcription and analysis in a single prompt. The transcription suffers; do it in two passes.
FAQ
- Can ChatGPT read handwriting?: Printing yes, neat cursive usually, messy cursive poorly. Surface uncertainty with [?] and verify critical lines yourself.
- Why did it invent a column that’s not in my chart?: Usually because the chart was small or the legend was cropped. Re-upload a cleaner crop and constrain the prompt to “only what is explicitly labeled.”
- Does Vision work on math equations?: Reasonably for printed equations; poorly for handwritten ones. Ask it to render the equation in LaTeX so you can compare it back to the source.
- Is it OK to upload screenshots with personal data?: Check your plan’s data-retention settings. As a rule, redact identifiers before uploading anything that wouldn’t be safe in a regular chat.