Claude Vision Prompts: Working with Images

Picture this: you upload a chart to Claude, type "what does this show," and get back something like "this appears to be a bar chart showing values across several categories." Technically true. Completely useless. That gap between a generic image description and an actually useful analysis is almost always a prompting problem, not a capability problem — claude vision prompts follow the same rule as text prompts: specific questions get specific answers, and vague questions get vague ones.

Before: The Weak Prompt

[Chart image attached]

What does this show?

What this does: it asks Claude to describe the image in general terms, which is exactly what you get back — a description of what type of visual it is, rather than an analysis of what the data actually means or why it matters.

Why It Fails

An open-ended question about an image invites an open-ended answer. Claude has no signal about whether you want the raw numbers, a trend analysis, a comparison to something else, or help spotting an anomaly — so it defaults to describing what's visually present, which is the safest but least useful response to a vague ask.

⚠️ Common mistake: treating an uploaded image like a magic box that will surface useful insight on its own. An image is data. It needs the same specific question a spreadsheet or document would need to produce a specific answer.

After: The Improved Prompt

[Chart image attached]

This is a chart of monthly signups for our product over the past 
year. Specifically:
1. What's the overall trend — growing, flat, or declining?
2. Identify any months that look like outliers compared to the 
   surrounding trend
3. Estimate the approximate values for the 3 highest and 3 lowest 
   months, based on the chart's axis
4. Note anything about the chart itself that makes it hard to read 
   accurately (unclear axis labels, overlapping data, etc.)

What this does: naming what the chart represents and asking four specific, answerable sub-questions turns a vague "what does this show" into a structured analysis Claude can actually work through point by point, rather than producing one generic paragraph.

⚡ Pro tip: always tell Claude what the image represents before asking your question, even when it seems obvious from the image itself. A chart labeled only with abbreviated axis names is often ambiguous without that context, and a wrong assumption about what's being measured can throw off everything downstream in the analysis.

Breaking Down Each Element

The fourth question — asking Claude to flag anything that makes the chart hard to read accurately — is the one people skip most often and the one that matters most for numeric precision. Vision-based reading of exact values from a chart is inherently less precise than reading them from a data table, and having Claude flag its own uncertainty (overlapping bars, an unclear legend, a compressed axis) tells you when to go find the underlying data rather than trust an estimated read.

A operations analyst at a retail company used this structure to review dozens of screenshotted dashboards from a legacy reporting tool that didn't export data cleanly. Asking explicitly for outlier months and flagged readability issues, rather than a general summary, let her catch two months where the chart's scale had visually compressed a real spike into something that looked unremarkable at first glance.

⚡ Pro tip: for anything where precision matters — financial data, medical charts, scientific figures — treat Claude's read of an image as a first-pass estimate to verify against the source data, not a final number to act on directly.

Variations for Different Contexts

For screenshots of text-heavy content (a webpage, a document scan, a slide), the equivalent structure shifts from "describe the trend" to "extract specific fields" — naming exactly which pieces of information you need pulled out (a date, a total, a specific clause) rather than asking Claude to summarize the whole image, which produces the same generic-description problem in text form.

For comparing multiple images — before/after photos, two versions of a design mockup — explicitly ask for a structured comparison across named dimensions (layout, color, specific element changes) rather than "what's different," which tends to surface only the most visually obvious change and miss smaller ones.

For photos of physical objects or spaces — a room layout, a piece of equipment, a whiteboard sketch — naming your actual goal ("I'm trying to figure out if this fits in a corner space that's roughly 6 feet wide") gets a more useful answer than describing the photo itself, since Claude can reason toward your goal instead of just narrating what's visible.

Save and Reuse This

The pattern holds across every image type: state what the image represents, ask specific sub-questions instead of one open-ended one, and request a flag on anything that limits reading precision. That structure works whether you're analyzing a chart, a screenshot, or a photo, with the specific sub-questions being the only thing that changes.

If you find yourself uploading the same type of image repeatedly — weekly dashboard screenshots, recurring chart formats — save the question structure in PromptABCD with the image type as a placeholder, so reviewing this week's dashboard starts from a prompt that's already proven to extract what you actually need, not a fresh "what does this show" every time.