Lesson 16. Prompts with images, PDFs, and files
Modern AI models work not only with text but also with images, PDFs, audio, and video. Most users still don't fully use this capability. Knowing how to properly compose prompts with attached files is a serious advantage.
Topic breakdown
Multimodal prompting means sending the model text together with other formats (image, document, audio). GPT-4o, Claude, and Gemini accept such requests.
Product photos, work documents, screenshots, contract PDFs, diagrams — these can be uploaded and analyzed. But just because the model can 'see' doesn't mean it sees everything correctly.
How you ask, what to focus on, and what not to assume — all of this affects the result. 'Analyze the image' works worse than a specific question.
Most models accept multiple images in one conversation. Number them and refer to each separately — this increases accuracy.
What you'll learn
- know how to write prompts with attached images and documents
- apply the correct structure of a multimodal prompt
- work with product photos, contracts, and design files
- know limitations: what the model might not see
Lesson plan
What is multimodality?
The ability to send images, PDFs, audio, or video along with text. GPT-4o, Claude, and Gemini support such requests.
Specific questions for images
Not 'analyze the image' but a specific question: 'What audience segment is this packaging designed for?' — this gives better results.
Working with documents and PDFs
Upload a contract, report, or presentation and ask a specific question. The model will read the entire document and extract the needed information.
What the model might not see
Very small text, low-quality images, handwritten notes — models sometimes miss these. Use the instruction 'if something is unclear — tell me'.
Weak vs strong prompt
Analyze this image.
You are a UX specialist in e-commerce. Based on this product card image: 1) identify the main UX issues; 2) give 3 recommendations for improving conversion; 3) evaluate colors and typography. Rely only on what you see.
The second prompt sets a role, a specific task list, and prevents guessing with 'rely only on what you see'. Result — structured and useful analysis.
Ready prompt template
Copy and adapt[Image or document attached]. You are a [role]. Based on this [image/document/file], do the following: 1) [first task]; 2) [second task]. Rely only on what you see or read — if something is unclear, tell me. Output format: [desired format].
Why it works
Attaching an image isn't enough. You need to clearly indicate what exactly the model should see and analyze.
For documents, formulate a specific question: 'What are the payment terms in this contract?' works better than 'Analyze the contract'.
The instruction 'rely only on what you see' protects against hallucination and increases reliability of the answer.
When working with multiple images, number them and address each: 'in the first image', 'in the second image'.
Practice
- Find an image or screenshot related to your work.
- Send it to ChatGPT or Claude with the request 'Analyze this image' and record the response.
- Send the same image with 3 specific questions.
- Compare both responses and evaluate the difference from specific formulation.
Mini-project
Mini-project: document or image analysis
Take a work document, contract, or product photo and conduct a full analysis with AI.
Tasks
- Choose a document or image for analysis.
- Formulate 3 specific questions: what exactly you need to learn.
- Upload to ChatGPT or Claude and ask those questions.
- Check the result: what did the model see correctly, what did it miss.
Deliverables
- analyzed document or image
- 3 questions and AI responses
- accuracy evaluation: what's correct, what's not
Checklist
Common mistakes
- generic request 'analyze the image' — the model doesn't know what's important
- expecting the model to see what it can't: tiny text, low-quality images
- not adding context to the image
- uploading many images without a specific question for each
Lesson FAQ
Can all models work with images?
No. GPT-4o, Claude 3, and Gemini Pro support images. Older models and text-only models cannot process images.
Can PDFs be uploaded directly?
Claude and Gemini accept PDFs directly. In ChatGPT, PDFs can be inserted as text or uploaded through Code Interpreter.