PROMPT ENGINEERING

Multimodal Prompts: Get More From Images and Files

Twenty new prompts for receipts, contracts, whiteboards, screenshots and product photos. They all start by making the model show you what it read.

·9 min read

Quick Answer

A multimodal prompt combines an image or file with written instructions. The most reliable pattern is image first, then a request that makes the model transcribe what it sees before analysing it, and mark anything unreadable instead of guessing. This week's 20 new prompts all follow that pattern, taking the library to 712 prompts.

The best way to prompt AI with a photo, scan or screenshot is to put the image first, ask the model to transcribe what it can see before it does anything else, and tell it to mark unreadable parts rather than fill them in. That one change fixes most of the bad results people get from image prompts.

It matters more than it used to. A lot of real work now starts with a picture rather than a paragraph: a receipt from a client lunch, a supplier quotation in Chinese, a whiteboard after a planning session, an error message on someone else's screen. Every major assistant will happily read these. The trouble is that it reads them confidently even when it is reading them wrong.

Twenty prompts went into the library this week to handle that. They cover all ten categories, two each, and every one is built around something you upload.

What Are the 20 New Prompts For?

They group into four kinds of job.

Reading documents you did not write. Rewrite a Scanned Official Letter in Plain English, Extract Key Terms From a Scanned Contract, Translate a Supplier Quotation From a Photo, and Turn a Textbook Page Photo Into Revision Cards. These are the high-stakes ones, where a misread date or amount has consequences.

Turning paper and screens into structured data. Turn Receipt Photos Into an Expense Claim Email, Turn a Handwritten Order Book Into a Spreadsheet, Turn Business Card Photos Into Follow-Up Messages, Turn a Whiteboard Photo Into Action Items, Plan My Week From a Calendar Screenshot, and Compare Competitor Prices From a Price List Photo.

Reviewing visual work. Review My CV Layout From a Screenshot, Audit My Instagram Grid From a Screenshot, Mark a Handwritten Worksheet From a Photo, Debug an Error From a Screenshot, Draft a Reply From an Email Thread Screenshot, and Turn a UI Mockup Image Into HTML and CSS.

Creating from your own photos. Write Captions From a Batch of Product Photos, Write Alt Text for Website Images From Photos, Restyle a Product Photo for a Festive Campaign, and Create a Professional Headshot From a Casual Photo.

The two regional ones are worth calling out. The supplier quotation prompt keeps the original language in a column beside the translation, so you can quote the supplier's own wording back to them. The order book prompt is for the many small businesses across the region that still take orders in a handwritten book, and it outputs CSV you can paste straight into a spreadsheet. Both sit in the SEA Business category.

Why Do Image Prompts Fail Differently From Text Prompts?

With a text prompt, the model gets exactly what you typed. With an image, it first has to read the image, and that reading step can go wrong without telling you.

The vendors are unusually direct about this. Anthropic's vision documentation says Claude "might hallucinate or make mistakes when interpreting low-quality, rotated, or very small images under 200 pixels." It also says counts of objects are approximate, that the model cannot reliably tell whether an image is AI-generated, and that heavy JPEG compression can make text difficult to read. Its advice ends with a plain instruction: always carefully review and verify image interpretations, especially for high-stakes use cases.

The result is a specific failure pattern. A misread digit on a receipt becomes a wrong total. A blurred clause in a contract gets filled in with something that sounds like a clause. A handwritten "7" becomes a "1". The answer looks exactly as polished as a correct one, which is the problem.

So the new prompts are designed around the reading step rather than the analysis step. Analysis is the part models are good at. Reading a phone photo taken at an angle under fluorescent light is the part that needs guarding.

Which Comes First, the Image or the Prompt?

First, when you have one image and a choice.

Anthropic's documentation says that, just as placing long documents before a query improves text prompts, Claude works best when images come before text, while noting that other placements still perform well. Google gives similar advice in its Gemini guide to prompting with media files: for a prompt with a single image, Gemini "might perform better if that image (or video) is placed before the text prompt." Both note that other placements still work, and Google adds that when images need to be interleaved with text to make sense, you should use whatever order is most natural.

In a chat app this simply means attaching the photo, then typing the prompt, rather than typing first and attaching at the end. It is a small habit, and it costs nothing.

Google's guide adds two further points that shaped this batch. Be specific: its example is an airport departures board, where "describe this image" gets a general description, but asking for the times and cities gets the data you wanted. And if the output is too generic, ask the model to describe the image before giving it the task. That second tip is the basis of the next section.

Why Should the Model Transcribe Before It Analyses?

Because it gives you a checkpoint. Almost every new prompt opens with a step like "transcribe the key sentences exactly as written" or "describe in one line what is actually visible, so I can check you read it correctly."

That step does two jobs. It forces the model to commit to what it saw before it starts reasoning, which Google's guide recommends for exactly this purpose: asking the model to describe the image shows you its initial understanding, so you can tell whether a bad answer came from bad reading or bad reasoning. And it gives you something quick to check against the photo. Comparing a ten-row table with a receipt takes thirty seconds. Discovering a misread amount after it has gone to your finance team takes longer.

The pattern shows up in different forms across the batch. The contract prompt asks for the clause wording and a page number for every row. The price list prompt transcribes the competitor's prices into a table before comparing anything. The caption prompt asks for a one-line description of each photo first, which catches the classic mistake of a pandan cake being read as matcha before it reaches five captions.

It is the same idea as asking a colleague to read back a phone number. Nobody thinks it is an insult, and it catches the error at the cheapest possible moment.

How Do You Stop the Model Guessing at Blurry Text?

Give it a word to use instead of a guess. Every prompt in this batch names one: UNCLEAR, UNREADABLE, ILLEGIBLE, NOT FOUND, or placeholders like OWNER? and DATE?.

This works because a model asked to complete a table will complete it. If you have not told it what to write when it cannot read something, plausible text is what it produces. Naming the alternative makes the honest answer as easy as the invented one. We covered the same principle for text in our guide to negative prompting: telling a model what not to do works best when you also tell it what to do instead.

Two further instructions do a lot of quiet work here. "Do not estimate an amount" and "do not convert currencies" appear in the receipt and quotation prompts, because those are exactly the places a helpful model will fill a gap with arithmetic you did not ask for. And the image editing prompts repeat "keep the product exactly as it is" in several ways, covering the label, colours, text and proportions, because image models are prone to tidying up things they were told to leave alone.

Photo quality still matters more than any instruction. Flat, straight on, in good light, with the whole page in frame. The pro tips on these prompts say so repeatedly because it is the single biggest factor in how often you see the word UNREADABLE.

What Should You Think About Before Uploading a Photo?

What is in the frame besides the thing you need. Photos carry more than you intend: a business card has a phone number, a worksheet has a child's name, a receipt can show part of a card number, and a whiteboard often has someone else's notes in the corner.

The prompts that handle personal information say so in their tips. The worksheet prompt suggests covering the student's name before you take the photo. The business card prompt is for cards people handed you to stay in touch, which is a different thing from uploading a contact list. And the contract prompt states plainly that it is a reading aid, not legal advice.

Beyond that, check your own plan's data settings and your company's policy before uploading anything sensitive. A cropped photo usually works just as well, and it is a smaller exposure if something goes wrong.

How Should You Use These This Week?

Pick the one that matches a pile on your desk. For most people that is receipts, business cards from an event, or a stack of photos for social media.

There is good reason to start small. Singapore's Singapore Digital Economy Report 2025 from IMDA found that AI adoption among SMEs tripled from 4.2% to 14.5% in 2024, driven mainly by uptake of off-the-shelf generative AI tools. Among larger firms it rose from 44.0% to 62.5%. Across the region, Google, Temasek and Bain's e-Conomy SEA 2025 report found consumer interest in AI topics running at three times the global average, and three out of four users said AI-powered tools had helped them discover content and make tasks easier.

In other words, most small businesses are only just starting, and the tools they are starting with are the same chat apps these prompts are written for. A prompt that turns a shoebox of receipts into a checked table is a better first step than an ambitious automation project.

If you are new to adding examples to a prompt, our few-shot prompting guide pairs well with this batch. Google's guide notes that a few realistic examples help with image tasks too, and the order book prompt's tip is exactly that: run one page first, fix the abbreviation list, then reuse it.

The library now holds 712 prompts across ten categories, all free and usable without an account. Last week's batch, on prompts for work with numbers in it, covers the analysis side of several of these jobs.

What Else Do People Ask?

Do these prompts work in ChatGPT, Claude and Gemini?

Yes. All three accept image uploads in their apps, and every prompt in this batch is written to be model-neutral. The limits differ by plan rather than by prompt. ChatGPT's pricing page, for example, lists uploads as limited on its Free plan, so if you plan to run a stack of receipts or business cards through it, check your plan's upload allowance first. WhichAIBest's ChatGPT review lists what the Free plan caps.

Should the image go before or after the prompt?

Before, when you have a single image and the choice is yours. Anthropic's documentation says Claude works best when images come before text, and Google's Gemini guidance says a single-image prompt may perform better with the image placed first. Both note that other orders still work, so when images and instructions need to be interleaved to make sense, use whatever order is most natural.

Why does the AI misread numbers in photos?

Usually because of the photo, not the model. Anthropic lists low-quality, rotated and very small images under 200 pixels as cases where Claude may hallucinate or make mistakes, and warns that heavy JPEG compression can make text hard to read. Photograph documents flat, straight on and in good light, and ask the model to transcribe figures before it uses them, so you can catch a misread digit early.

Is it safe to upload receipts, contracts and business cards?

It depends on your plan's data settings and your company's policy, so check both before you upload. A practical rule is to share only what the task needs. Cover student names on worksheets, crop out bank details that are not relevant, and avoid uploading other people's personal data in bulk unless you are permitted to. The prompts work just as well on a cropped or partly covered image.

Can AI edit my product photo without changing the product?

Often, but not reliably, which is why the festive campaign prompt repeats the instruction to keep the product exactly as it is, including the label and proportions. Always check the label and any lettering at full size before you post. If anything has changed, ask again with the same instruction rather than trying to fix the edit by hand.

Sources: Anthropic, Vision documentation, for image-before-text placement, the listed limitations on low-quality, rotated and small images, counting and AI-generated image detection, and the note on JPEG compression. Google, File prompting strategies for the Gemini API, for image-first ordering, specificity, few-shot examples, and describing the image before the task. IMDA, Singapore Digital Economy Report 2025, for SME and non-SME AI adoption. Google, Temasek and Bain, e-Conomy SEA 2025, for regional AI interest and usage. OpenAI, ChatGPT pricing, for upload limits on the Free plan.