Most people use Claude as a text tool. But Claude can see. Upload a screenshot, a photo, a diagram, a chart, or a document scan — and Claude can read, analyze, describe, extract, and reason about what it sees. Combine that with text instructions, and an entirely new category of tasks becomes possible. Error messages that were confusing become explained. Data trapped in image-based PDFs becomes extractable. Wireframes become code. Visual information that previously required manual transcription, multiple tools, or simply couldn't be processed now flows directly into Claude's reasoning.
Multimodal prompting combines visual inputs with text instructions to accomplish tasks neither alone can do. Claude can read text in images, interpret charts and diagrams, analyze UI layouts, and reason about spatial relationships. The chapter covers three layers: what Claude sees well and where its limits are, six analysis patterns (screenshot explanation, data extraction, document digitization, design review, competitive analysis, chart analysis), four multimodal workflows (screenshot to action, document to insight, visual to code, image comparison), domain-specific use cases for five roles, and five best practices — with context provision being the single most impactful.
What Can Claude See — and Where Are the Limits?
Understanding what Claude actually processes when it sees an image prevents both under-use (not knowing to try) and over-reliance (trusting every value extracted from a low-resolution chart without verification).
- Text in images — Screenshots, scanned documents, handwritten notes, text on photos, image-based PDFs
- Visual structure — Charts and graphs, tables, diagrams, flowcharts, UI layouts and wireframes
- Visual content — Photos, illustrations, logos, physical objects and environments, code in screenshots
- Spatial relationships — Layout and positioning, relationships between diagram components, before/after comparisons
- Very small text — Fine print at low resolution may be misread
- Heavily stylized fonts — Artistic or decorative text can be misinterpreted
- Low contrast images — Poor visibility affects accuracy
- Highly complex diagrams — Very dense technical drawings may be partially misread
- Real-time input — Static images only, not live video or screen sharing
- Precise measurements — Claude approximates, not measures exactly
Always verify critical information extracted from images against the source — especially numbers from charts or data grids used in financial or operational decisions.
How Do You Analyze Screenshots and Documents?
This is the most immediate use case for most professionals — processing visual information that would otherwise require manual reading, copy-pasting, or switching between tools.
Pattern 1: Screenshot Explanation
Use cases: error messages you don't understand, unfamiliar software interfaces, analytics dashboards you've inherited, reports from tools you don't use daily.
Pattern 2: Data Extraction from Images
When data is trapped in an image and copy-paste doesn't work — charts from PDFs, scanned spreadsheets, tables in screenshots — this pattern extracts it cleanly while flagging uncertainty.
Pattern 3: Document Digitization
Use cases: handwritten meeting notes, whiteboard photos from sessions, physical documents that need digitizing, old printed materials.
How Do You Use Claude for Visual Analysis and Critique?
Beyond reading, Claude can evaluate, critique, and provide expert analysis of visual content — assessing design decisions, extracting competitive intelligence, and interpreting data visualizations.
Pattern 4: Design Review
Pattern 5: Competitive Visual Analysis
Pattern 6: Chart and Graph Analysis
What Are the Four Multimodal Workflow Patterns?
The real power of vision emerges when it is combined with Claude's other capabilities in structured workflows. Each of the four patterns below represents a class of task that previously required manual steps, multiple tools, or significant time.
Workflow 1: Screenshot to Action
See something, understand it, take action. The most common multimodal workflow.
Real example: Upload screenshot of a slow webpage → "What performance issues can you identify?" → "Write the CSS/JavaScript fixes for the issues you identified."
Workflow 2: Document to Insight
Transform visual documents into actionable insight without manual transcription.
Real example: Upload whiteboard photo from strategy session → "Transcribe and organize by theme." → "Draft a follow-up email summarizing the session decisions."
Workflow 3: Visual to Code
Particularly powerful for developers and designers — describe a wireframe or UI in natural language, then build it.
Real example: Upload wireframe sketch → "Build this as a responsive HTML/CSS layout" → "Add navigation hover states and make the header sticky."
Workflow 4: Image Comparison
Use cases: A/B test variants, before/after design comparisons, version comparisons of documents or interfaces, brand consistency checks.
How Do Different Roles Use Vision Capabilities?
The same underlying capabilities take very different forms depending on who is using them and what they are trying to accomplish. The prompts below provide role-specific starting points.
- Visual hierarchy effectiveness
- Typography choices and readability
- Color harmony and contrast ratios
- Whitespace and breathing room
- Alignment and grid consistency
- Brand consistency checks
- Diagnose bugs from screenshots
- Build UI from wireframes
- Explain code shown in images
- Identify likely error causes
- Generate fix code with verification steps
- Extract all quantitative values visible
- Assess performance relative to benchmarks
- Write 2-paragraph executive summaries
- Recommend 3 most important actions
- Flag anomalies and trends
- Plain-language explanation of figures
- Walk-through of axes, scales, and legend
- Key insight of each visualization
- Why it matters for the topic
- Comparison with other figures
- Style analysis of reference images
- What makes competitor content effective
- Gaps and differentiation opportunities
- Adaptation guidance: "in this style but distinctively mine"
The Domain Grid above shows something important: the same underlying capability — "analyze this image" — produces completely different outputs depending on the context and role supplied. Claude's vision does not have a single mode. The role context in the prompt determines which analytical lens gets applied to the visual information.
What Are the Five Best Practices for Vision Prompting?
Practice 1: Provide Context Before Analysis
[Upload image] "What do you think?"
[Upload image] "I'm a UX designer reviewing our onboarding flow. This is the third screen new users see. Please analyze for clarity, visual hierarchy, and any friction points."
Context shapes what Claude looks for and how it interprets ambiguity. The same image will produce very different analysis under "I'm a marketer reviewing competitor positioning" versus "I'm a developer debugging a UI issue."
Practice 2: Be Specific About What to Extract
"Extract the data from this chart."
"Extract the monthly revenue figures for 2023. Present as a table with Month and Revenue columns. Flag any months where you're uncertain about the value."
Practice 3: Ask for Uncertainty Flags
This is especially important when extracted data will be used in decisions. Claude will often identify its own uncertainty correctly when explicitly asked — but it will not flag uncertainty by default unless prompted.
Practice 4: Use Sequential Analysis for Complex Images
For complex diagrams or dense documents, breaking the analysis into steps produces far more accurate results than asking for everything at once.
Practice 5: Combine Images with Text Context
"Here's the chart from our Q3 report [image]. The context is that we launched a new pricing model in August. Analyze how the pricing change appears to have affected the metrics shown."
The image gives Claude the visual data. The text gives Claude the interpretive context. Together they produce analysis that neither could generate alone.
What Are the Most Common Vision Prompting Mistakes?
Upload image, ask "what is this?" — produces a generic description with no analytical value.
Explain your role, your goal, and exactly what you need from the image before asking for analysis.
"What is the exact revenue figure for March?" when the chart is low-resolution.
"Estimate the March revenue figure and note your confidence level. Flag if the resolution makes this uncertain."
Using Claude's data extraction from images in financial reports without checking against the source.
Always verify extracted numbers against the original source for any high-stakes use.
Assuming every character in every image is read correctly, especially with handwriting or stylized fonts.
Ask for uncertainty flags — Claude will identify its own doubt correctly when prompted to do so.
Asking for ten different analyses of a complex diagram in a single prompt.
Sequential analysis — one aspect at a time for complex visuals — produces far more accurate and useful results.
- Claude sees text, structure, and visual content — Screenshots, documents, charts, diagrams, photos
- Context transforms analysis — Always explain who you are and what you need before asking
- Request uncertainty flags — Especially for critical data extraction from images
- Multimodal workflows multiply value — Vision + text + action unlocks new capabilities
- Sequential works better for complexity — Break dense images into focused questions
- Verify extracted data — Especially numbers from low-resolution or complex images
- Vision removes manual transcription — What used to require copy-pasting now just works
Challenge: Use vision capabilities on a real task this week.
Beginner Option
Upload a screenshot of an interface that feels confusing — an app, dashboard, or error message. Ask Claude to explain what is visible and what to do.
Intermediate Option
Upload a chart or report image from your work. Ask Claude to extract the key data, identify the trends, and write a two-paragraph executive summary.
Advanced Option
Run a full multimodal workflow: upload a visual input, extract or analyze it, then use the output to generate something — code, copy, a plan, or a recommendation.
Reflection Questions
What information was Claude able to extract that would have been tedious to do manually? Where did Claude flag uncertainty — and was it right to be uncertain? What multimodal workflow could save the most time in your regular work?
In Chapter 20: Marketing & Sales Enablement, the series turns to campaign development, sales collateral, positioning frameworks, competitive analysis, and funnel optimization.
Claude can read and analyze text in images including screenshots, scanned documents, handwritten notes, and text on photos. It can interpret visual structure such as charts, graphs, tables, diagrams, flowcharts, and UI layouts. It can describe visual content including photos, illustrations, logos, and physical objects, and reason about spatial relationships and layout. Limitations include very small or low-resolution text, heavily stylized fonts, low-contrast images, highly complex technical diagrams, and precise measurements — Claude approximates rather than measures exactly.
Multimodal prompting is the practice of combining visual inputs with text instructions to accomplish tasks that neither alone can do. Uploading a screenshot of a confusing error message and asking for an explanation, extracting data from a chart image where copy-paste does not work, or uploading a wireframe and asking Claude to build it as HTML are all multimodal workflows. These tasks previously required multiple tools, manual transcription, or were not possible at all. The key principle is that the image provides visual data while the text provides interpretive context.
Specify the exact format you need — table, JSON, CSV, or list — and list the specific fields to extract. Always ask Claude to flag values it is uncertain about after extracting. For low-resolution images or complex charts, acknowledge the limitation and ask for a confidence level alongside each value. For critical data, always verify extracted numbers against the original source — Claude approximates from visual patterns and cannot guarantee character-level accuracy for every value.
The four patterns are: Screenshot to Action — see a problem in a screenshot, understand it, then generate the fix; Document to Insight — transcribe a visual document, extract key information, then synthesize a response or summary; Visual to Code — describe a wireframe or UI screenshot, then build it as HTML, React, or another format; and Image Comparison — upload two images side by side to identify differences, assess which is more effective, and get recommendations on what to keep or change.
Providing context before asking for analysis is the single most impactful practice. Uploading an image and asking "what do you think?" produces a generic description. Explaining your role, what the image represents, and exactly what you need — "I am a UX designer reviewing our onboarding flow; this is the third screen new users see; please analyze for clarity, visual hierarchy, and friction points" — produces targeted, actionable analysis. Context shapes what Claude looks for and how it interprets ambiguity in the image.