10 Claude Prompts for Image Analysis That See What You Miss

Claude can see. When you attach an image to a conversation, Claude does not just acknowledge it — it can read text within it, describe what is happening in it, identify objects and people, interpret charts and diagrams, spot problems in designs, and extract structured data from visual content. But like every other interaction with Claude, the quality of the output depends almost entirely on the quality of the prompt. The right Claude prompts for image analysis tell Claude exactly what to look for, how to structure its response, and what level of detail you need — turning a capable vision model into a precise analytical tool.

Below are 10 prompt patterns that cover the most valuable image analysis use cases — from extracting text to interpreting data visualisations to reviewing UI designs. Each includes a ready-to-use example, an explanation of why it works, and a tip for getting even more from it.

Claude's vision capability is genuinely powerful, but it defaults to general description when given no specific instruction. “Describe this image” produces a general caption. “List every piece of text visible in this image, formatted as a numbered list” produces a precise, structured inventory. The difference between those two prompts is the difference between an observation and an analysis.

The prompts below are built around that principle. They are designed to give Claude a specific task, a structured output format, and enough context about what you are trying to accomplish to produce output that is immediately useful rather than requiring further interpretation or reformatting.

Extracting text from images — screenshots, scanned documents, photos of whiteboards, slides, signage — is one of the highest-frequency image tasks in everyday work. This prompt extracts all visible text accurately and organises it in a way that is immediately usable.

"Extract all text visible in this image. Preserve the original formatting and hierarchy as closely as possible: headings should appear before body text, bullet points should remain as bullets, columns should be separated clearly. If any text is partially obscured or unclear, flag it with [unclear] rather than guessing. Output the extracted text only — no commentary."

Why it works: The instruction to flag unclear text rather than guess is critical — it prevents Claude from confidently outputting incorrect text that you then rely on. Preserving hierarchy means the extracted text is immediately usable rather than a flat blob that needs manual restructuring. The no-commentary instruction keeps the output clean for downstream use in documents or databases.

Sometimes you need a thorough, methodical description of everything in an image — for accessibility alt text, for cataloguing visual assets, for generating training data, or for communicating about an image to someone who cannot see it. This prompt produces a structured, comprehensive description.

"Describe this image in detail, structured as follows: (1) Overall scene — one sentence summary of what the image shows, (2) Main subjects — who or what is the primary focus, described specifically, (3) Background and setting — environment, location, time of day if visible, (4) Notable objects — any items, text, logos, or symbols present, (5) Visual style — photography, illustration, screenshot, diagram, etc. and any distinctive aesthetic qualities. Be specific and objective — describe what is visible, not what you infer."

Why it works: The five-part structure ensures systematic coverage rather than a description that focuses on what catches the eye and ignores the periphery. The instruction to describe what is visible rather than what is inferred prevents Claude from adding interpretive commentary that may be incorrect. This structure is particularly well-suited to generating alt text that is both descriptive and appropriately detailed.

Charts, graphs, and data visualisations are only as useful as your ability to read them quickly and accurately. This prompt extracts the key insights from any visualisation — bar chart, line graph, pie chart, scatter plot, dashboard — and presents them as clear, actionable findings.

"Analyse this data visualisation. Provide: (1) Chart type and what it is measuring, (2) The key data points — exact values where legible, approximate where not, (3) The top 3 insights or trends visible in the data, stated as assertions not observations, (4) Any anomalies, outliers, or data points that stand out from the pattern, (5) What is missing or unclear that would be needed to fully interpret this data. Do not describe the visual appearance of the chart — focus on what the data shows."

Why it works: The instruction to state insights as assertions rather than observations is the key distinction — “revenue declined in Q3” is an observation; “Q3 revenue was the lowest in 18 months, breaking a consistent growth trend” is an insight. The missing-information question is particularly valuable for catching charts that lack context — no baseline, no time range, no sample size — that would change the interpretation significantly.

Getting a fast, structured critique of a UI design, landing page, or marketing asset before it ships saves expensive revision cycles later. This prompt reviews a design image across the dimensions that actually affect user experience and conversion.

"Review this UI design / webpage / marketing asset as an experienced UX designer. Evaluate it on: (1) Visual hierarchy — does the eye flow naturally to the most important elements, (2) Clarity of the primary call to action — is it obvious what the user should do next, (3) Readability — font sizes, contrast ratios, line length, (4) Consistency — are spacing, colours, and typography consistent throughout, (5) Any accessibility concerns visible at a glance. For each category, identify the single most important issue and suggest a specific fix. Do not give generic design advice — reference specific elements visible in this image."

Why it works: Asking for the single most important issue per category forces prioritisation rather than a laundry list of every possible improvement. The instruction to reference specific elements visible in the image prevents generic feedback. This is particularly useful as a fast pre-review before sharing a design with a client or stakeholder who will have limited patience for obvious problems.

For e-commerce cataloguing, inventory management, document processing, and asset tagging, you often need a structured inventory of everything visible in an image rather than a narrative description. This prompt produces a clean, structured list.

"List every distinct object, item, or element visible in this image as a structured inventory. For each item include: name, quantity if more than one, colour, approximate position in the image (top-left, centre, background, etc.), and any visible text or branding on it. Format as a table with columns: Item | Quantity | Colour | Position | Text/Branding. If you cannot determine an attribute with confidence, write 'unclear' rather than guessing."

Why it works: The table format makes the output immediately usable for database import, spreadsheet analysis, or structured reporting without reformatting. Position information is often overlooked in image descriptions but is critical for spatial reasoning tasks. The “unclear rather than guessing” instruction maintains data quality for downstream use cases where a confident wrong answer is worse than an honest unknown.

Processing scanned forms, receipts, invoices, business cards, and official documents is one of the highest-value image processing tasks in business workflows. This prompt extracts structured data from document images in a format ready for downstream use.

"Extract all structured data from this document image. Output as JSON with the following fields where present: document_type, date, issuer (name and address), recipient (name and address), line_items (array of item/description, quantity, unit_price, total), subtotal, tax, total_amount, reference_number, payment_terms. For any field not present in the document, omit it from the JSON. Flag any values you are not confident about with a note field: {value: '...', confidence: 'low', note: '...'}."

Why it works: JSON output is directly consumable by any downstream system — databases, APIs, spreadsheet imports, automation workflows. The confidence flag for uncertain values is essential for any automated processing pipeline where a confidently wrong value would cause downstream errors. Omitting absent fields rather than returning null keeps the output clean and schema-flexible.

Comparing two images — a before and after, two design options, two product variants, two versions of a document — requires systematic side-by-side analysis. This prompt structures a comparison that highlights differences clearly rather than describing each image independently.

"Compare these two images systematically. For each dimension below, describe what is the same and what is different: (1) Content — what is present in one but not the other, (2) Layout and composition — how elements are arranged, (3) Colour and visual style, (4) Text — any differences in copy, labels, or headings, (5) Overall impression — which is clearer, more polished, or more effective and why. Format your response as a comparison table for the first four dimensions, then a paragraph for the overall impression."

Why it works: Structuring the comparison by dimension rather than image-by-image prevents the response from becoming two independent descriptions with no comparative insight. The table format makes differences scannable at a glance. The overall impression paragraph is where the most useful judgement lives — and separating it from the table prevents it from contaminating the objective comparison with subjective assessment.

Technical diagrams — system architecture, process flows, network maps, org charts, circuit diagrams — contain dense information that can take significant time to parse. This prompt translates a visual diagram into a clear verbal explanation that captures both the structure and the logic.

"Explain this diagram clearly. Structure your explanation as: (1) What type of diagram this is and what system or process it represents, (2) The main components or nodes — what each one is and what it does, (3) The connections and flow — how components relate to or communicate with each other, (4) The overall sequence or logic — how the system works from start to finish or from input to output, (5) Any notable features, bottlenecks, decision points, or dependencies that stand out. Write for an audience who understands the domain but has not seen this specific diagram before."

Why it works: The five-part structure moves from identification to components to connections to logic — the natural reading order for understanding any diagram. The audience specification prevents Claude from either over-explaining basics or under-explaining domain context. The bottlenecks and dependencies question is particularly valuable for technical diagrams where the most important insight is not the flow itself but the points where the flow is most likely to fail.

If you have an image with a visual style you want to recreate in an AI image generator, this prompt reverse-engineers the image into a precise generation prompt — capturing subject, composition, lighting, style, colour palette, and mood in the format image generation tools understand best.

"Analyse this image and write a detailed prompt I could use to recreate a similar image in an AI image generator like Midjourney or DALL-E. The prompt should capture: subject and action, composition and framing, lighting style and direction, colour palette and mood, photographic or artistic style, any distinctive visual techniques — bokeh, grain, high contrast, illustration style, etc. Write the prompt in the format these tools respond to best: specific, descriptive, and comma-separated. Also provide 3 negative prompt suggestions for things to exclude."

Why it works: Most image generation prompts fail because they describe the subject but not the visual treatment. This prompt forces Claude to extract the full set of visual qualities — not just what is in the image but how it was shot or rendered. The negative prompt suggestions are valuable for excluding the specific artefacts or styles that would undermine the recreation — often the most important part of getting consistent results from image generators.

Writing accurate, appropriately detailed alt text for every image is one of the most consistently neglected accessibility tasks in web and content publishing. This prompt generates compliant, context-aware alt text that serves both accessibility needs and SEO requirements.

"Write alt text for this image for use on a [describe the context: e.g. 'product page on an e-commerce website' / 'blog post about remote work' / 'news article about climate change']. Follow these rules: (1) describe what is actually in the image, not what you interpret it to mean, (2) include any text visible in the image, (3) keep it under 125 characters unless necessary, (4) do not start with 'Image of' or 'Photo of', (5) if this image is purely decorative and conveys no information, output only: alt=''. Also provide a longer description (2-3 sentences) for use in a figcaption if needed."

Why it works: The context specification changes what information is most relevant to include — alt text for a product image needs to describe the product specifically, while alt text for an illustrative blog image needs to describe what is happening. The 125-character guideline matches screen reader best practices. The decorative image rule prevents unnecessary noise for screen reader users. The figcaption option gives you a longer format for images where more context genuinely helps.

The single most important factor in image analysis quality is specificity about what you want extracted and how you want it formatted. Claude will default to a general description if given no specific instruction. Every prompt above replaces that default with a precise task, a structured output format, and rules for handling uncertainty. That combination — task + format + uncertainty handling — is the pattern that produces consistently useful results across any type of image.

Save the prompts that match your most common image processing tasks as reusable templates in Chat Smith so you can deploy them instantly — the Text Extractor for screenshots, the Document Extractor for invoices, the Chart Interpreter for analytics dashboards. Each prompt becomes more valuable the more precisely you know when to reach for it.

Using these prompts steers you away from the most consistent image analysis failures. Unstructured description requests produce general captions that bury the specific detail you needed. No uncertainty instructions produce confidently wrong text extractions that corrupt downstream data. No format specification produces narrative output that requires manual reformatting before it can be used. No context for the intended use produces alt text that is either too vague or too verbose for the specific application.

Each prompt in this guide addresses one of these failure modes directly. The Text Extractor addresses confident hallucination with the unclear flag. The Chart Interpreter addresses vague observation with the assertion instruction. The Document Extractor addresses unusable output with JSON formatting. The Alt Text Generator addresses context-free description with the placement specification. The pattern is always the same: the more precisely you specify the task, the more precisely Claude executes it.

Images contain far more information than most workflows ever extract from them. These 10 Claude prompts for image analysis give you a systematic way to unlock that information — from structured data extraction to design critique to diagram interpretation — for any image, in any context. Start with the prompt that matches your most frequent image task. Build the habit of specifying task, format, and uncertainty handling in every image prompt. The difference in output quality will be immediate.

Image analysis workflows often involve repeating the same type of extraction or interpretation across many images — processing a batch of invoices, reviewing a series of design iterations, generating alt text for a content library. Keeping your best-performing image prompts organised and instantly deployable is exactly where Chat Smith comes in. Chat Smith is an all-in-one AI platform that lets you save every image analysis prompt as a reusable template, organise them by image type or use case, and launch any prompt in one click across Claude, GPT, Gemini, and other leading models.

Instead of rebuilding your invoice extraction prompt every time a new document comes in, or hunting for your chart interpretation template before a data review meeting, Chat Smith gives you a clean, searchable library of your best-performing prompts. You can run the same image prompt across multiple models to compare outputs, share your prompt library with teammates who handle the same image types, and build a consistent image processing workflow that scales with your volume.

1. What image formats does Claude support?

Claude supports JPEG, PNG, GIF, and WebP image formats. For best results with text extraction, use the highest resolution version of the image available — low-resolution images will reduce OCR accuracy. For charts and diagrams, screenshots at 2x or retina resolution produce noticeably better results than compressed or downscaled versions.

2. How accurate is Claude at reading text in images?

Claude performs well on clear, high-contrast printed text and typed text in screenshots. It performs less reliably on handwritten text, very small text, text on complex backgrounds, or text at an angle. The Text Extractor prompt’s instruction to flag unclear text rather than guess is specifically designed to handle these cases gracefully — an [unclear] marker is far more useful than a plausible-sounding wrong word.

3. Can I use these prompts to process multiple images at once?

Yes — Claude can process multiple images in a single message. For batch processing tasks like comparing several design options or extracting data from multiple documents, attach all images in one message and adjust the prompt to reference them by number or label. For very large batches, processing in groups of 3–5 images per message tends to produce more consistent and detailed output than processing all at once.

4. Can Claude identify specific people, logos, or brands in images?

Claude can identify many well-known logos, brand visual identities, and product designs. It does not identify specific individuals from their appearance as a matter of privacy policy. For logo and brand identification tasks, the Object and Label Inventory prompt is the most effective format — it captures the text/branding column for each identified item, which is where logo recognition most reliably surfaces.