Analyze an image by stating the task, supplying a clear original, defining what may be inferred, and asking for location-based evidence. Verify small text, counts, measurements, identity, and ambiguous details separately. Image-capable models can describe patterns, but they should not be treated as authoritative sensors or expert reviewers.
Who this is for: Designers, support teams, researchers, and operators using screenshots or photographs as evidence for a practical task.
- Image quality and framing determine what evidence is available before prompting begins.
- Separate visible observations from interpretations and recommendations in the response.
- Crop, zoom, or use specialized tools to verify details that carry operational risk.
Prepare the visual evidence
Use the highest practical resolution and avoid recompressed screenshots when exact text matters. Include enough surrounding interface or scene context to interpret the target, but remove unrelated personal or confidential information. For multiple images, label each one and state their order, date, and relationship.
Check orientation, glare, blur, crop boundaries, color distortion, and hidden overlays. A screenshot only captures one state and may omit off-screen content, hover text, or prior steps. Record these limitations so a reviewer does not mistake the image for a complete account of an event.
Ask observation before interpretation
Begin with a bounded request such as transcribing visible error text, listing interface elements, or identifying differences between labeled regions. Ask the model to separate directly visible details from inferences. Location descriptions such as upper-right panel or row three make claims easier to inspect.
Then request interpretation tied to those observations. For usability review, the model might connect a disabled control and missing explanation to a likely user obstacle. It should not invent the user's intent or claim that every user will respond the same way. Alternative explanations help prevent premature diagnosis.
Verify fragile visual details
Small text, dense tables, exact counts, subtle color differences, and partially obscured objects are common failure points. Crop the relevant region and ask again, but compare against the original. Use OCR, pixel measurement, logs, metadata, or domain instruments when those are the appropriate source of truth.
Do not use general visual analysis to identify a person, diagnose a medical condition, determine authenticity, or make a safety-critical judgment without appropriate methods and qualified review. The model's explanation can be plausible even when the visual premise is mistaken.
Test a repeated image workflow
Build a labeled set covering normal images, low resolution, unusual layouts, missing regions, and confusing distractors. Score each required observation separately. If the workflow extracts interface errors, test exact wording, location, and whether an error is present, rather than assigning one broad quality score.
Provider support for image formats, resolution handling, and request structure changes over time. Consult current official documentation and re-test after model changes. Retain human review for output that affects customers, safety, access, or records, and protect image privacy throughout storage and logging.
Triage a checkout error screenshot
Customer support receives a screenshot showing a failed purchase but no written description of the steps taken.
- Remove visible personal and payment details or obtain an appropriately redacted copy before analysis.
- Ask for exact visible error text, page state, selected options, and element locations without guessing a cause.
- Crop the error region and verify the transcription manually against the original screenshot.
- Compare observations with application logs and known incident reports for the same time and flow.
- Draft troubleshooting questions that distinguish missing context, then have a support agent decide the response.
Visual analysis request card
Attach this information when an image is used as input to a repeatable review.
- Task and boundary: exact observations needed, prohibited inferences, and intended decision.
- Image record: label, source, time, original resolution, redactions, and known missing context.
- Evidence format: observation, image label, region, transcription, and uncertainty note.
- Verification path: crop, OCR, log, metadata, measurement, or qualified reviewer for each fragile detail.
- Privacy handling: consent, minimization, retention, access, and safe logging requirements.
Common mistakes
- Uploading a tiny compressed screenshot and trusting exact text or numerical readings from it.
- Asking what happened without separating visible evidence from a guessed sequence of events.
- Including private background content that has no role in the analysis task.
Try one
A model says a dashboard screenshot proves sales fell because of a campaign. Evaluate that response and propose the next check.
The screenshot may show a visible sales decline and perhaps a campaign label, but it cannot by itself establish causation. A good answer records the observed axes, dates, values, and annotations, verifies the chart reading, then examines underlying data and alternative causes. Evaluation should reject causal certainty based only on visual co-occurrence.
Sources
- OpenAI images and vision guideOfficial guidance for providing images as model input and interpreting visual content.
- Gemini API image understanding guideOfficial Google guidance for interpreting image inputs.