Give AI a screenshot and ask the right question
The expenses app has refused your claim with an error code and a half-visible sidebar. You screenshot it and ask AI what's wrong, and it answers with great confidence about the bit you cropped off.
What you'll make
a short answer that lists what the screenshot shows, separates it from what the AI is guessing, and names the next check that would settle the question.
What you'll need
ChatGPT, Claude, Gemini or Copilot in your work account with image upload, a screenshot you’re allowed to share, and about ten minutes.
Example result.
Visible: "Error 403: Approval chain missing", on the Submit claim screen, claim £86.40.
Not visible: the left sidebar is cut off. No approver name appears.
Possible explanations: no approver set on your profile, or the cost centre has none.
Next check: open your profile and see whether an approver is listed.1. Take a screenshot that answers the question
Decide what you need to know before you capture anything. Include the whole message, the screen title, any field the message refers to, and units or labels on figures. A crop that removes the heading can make the message impossible to interpret.
Then check everything else in the frame: names, email addresses, other tabs, chat windows and notifications. Crop them out or use a cleaner screenshot. Don’t rely on a quick blur to hide sensitive text.
Why this works: the AI can only reason about what’s in the image. A good crop keeps the evidence and leaves out what you shouldn’t share.
What you should see: a screenshot with the message, its context and nothing personal you didn’t mean to include.
2. Upload it and ask for the visible evidence first
Attach, paste or drag the image into the chat, check the thumbnail is the right one, then send:
Read the attached screenshot and answer this question: [specific question].
First list the visible text, labels and values relevant to the answer, quoted exactly. Then separate what you can see from anything you infer. Mark anything unreadable or cut off as uncertain. Don't describe parts of the screen that aren't in the image, and don't state a cause as confirmed.[specific question] is one the image can support. “What does this error message say, and on which screen?” works. “Why is the system broken?” needs evidence outside the screenshot.
Why this works: asking for the visible evidence first means a misreading shows up before it turns into an explanation.
What you should see: quoted text from the screenshot, then a clearly separate section for interpretation.
3. For an error message, ask for the next check
An error message is evidence, not a diagnosis. Tell the AI what you were trying to do, then ask it to narrow the possibilities:
Here is the exact error and what I was trying to do: [task]. List up to three possible explanations, each tied to something visible in the screenshot or in what I've told you. Then give the single next check that would tell them apart. Do not state a cause as confirmed.[task] is a plain description, such as “submitting a travel expense claim for £86.40”.
Why this works: “the next check” turns a guess into something you can test in a minute, instead of a confident story you act on.
What you should see: two or three possibilities and one specific thing to look at or try.
4. Check the reading against the original
Open the screenshot at a useful size and compare. Check the exact wording, error codes, figures, decimal points and minus signs. If the answer quotes text you can’t find in the image, it came from somewhere else. If a figure matters for a calculation, use the underlying data, not a reading from a picture.
Why this works: small text, low resolution and similar-looking characters are where image reading slips.
What you should see: every quoted item matching the screenshot exactly.
Check the result
- Every quote in the answer appears in the screenshot, word for word.
- Anything cut off or unreadable is marked as unknown, not filled in.
- No cause is presented as confirmed.
- The next check is something you can actually do.
The usual failure: the answer describes the part you cropped off, or turns one possibility into the cause. Spot it by asking of each sentence: “Where in the image is this?”
The screenshot doesn't show [claimed detail]. Re-answer using only what is visible. Keep the possible explanations as possibilities, and give the single next check that would tell them apart.[claimed detail] is the part of the answer you can’t find in the image.
Worked example: an expenses error
Fictional example. Sam is submitting a travel claim and gets this screen. The left sidebar is partly cut off.
Screen title: Submit claim
Claim: Travel · £86.40 · 7 October 2026
Dialog: "Error 403: Approval chain missing. Contact your administrator."
Left edge: a sidebar, cropped, showing "Pro…" and "Cost c…"The first answer said:
Your profile has no manager assigned, as the sidebar shows. Ask IT to reset your permissions and resubmit.
- The failure. The sidebar is cropped. Nothing in the image shows a manager field, and “reset your permissions” is a guess at the fix.
- The repair. The correction prompt, with “the sidebar shows no manager assigned” as the claimed detail, produced the answer in the example result: the quoted error, the cut-off sidebar marked as not visible, two possibilities and one check.
- What happened next. Sam opened the profile page. An approver was listed, so the next check was the cost centre, which had none. One message to finance fixed it. IT never needed to hear about it.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Optional training: ask what's missing from a dashboard crop
For a busy dashboard, ask which labels, filters or sections are unreadable or outside the frame before asking what it means. A crop isn’t evidence that the missing part contains nothing important. If you need exact arithmetic, paste the underlying table instead. A screenshot can locate a figure, but a table is a better source for calculating with it.
Before answering, list any labels, filters, date ranges or sections in this screenshot that are unreadable or cut off. Then say which of them would change the answer to [question].Optional training: use a screenshot to find a setting
When you’re lost in an app, a screenshot of the current screen plus “where do I change X?” can work well. Ask for the steps in terms of the labels visible in the image. Then check the result in the app itself. A description of a button is not evidence that anyone clicked it.
Optional training: describe an image for someone else
Describing an image so another reader can use it is a different job from asking a question about it. Write useful image alt text with AI covers it, and W3C’s images tutorial explains the principles.