Documents, data and images

Some of the biggest time savings come from working with your own material — reports, contracts, spreadsheets, screenshots. Here's how to prompt for accurate results.

4.1 Working with long documents

Modern assistants can read very long documents, but how you ask still matters:

• Put long documents first and your question last. Anthropic's guidance notes that placing the query after the long material can noticeably improve response quality.
• Label each document — with tags or headings, including its title, date and source — so the model can refer to them.
• Ask for quotes first. "First extract the passages relevant to X, then answer using only those passages." This grounds the answer and makes checking easy.
• Ask for page or section references so you can verify.

<document title="Supplier contract — Acme Ltd" date="2026-03-01">
[contract text]
</document>

1. Quote every clause about termination, word for word, in <quotes> tags.
2. Then, using only those quotes, explain in plain English how either
   party can end the contract and what notice is required.
3. If something is ambiguous, say so rather than guessing.

4.2 Working with data

AI tools can analyse spreadsheets and CSV files — many can now write and run code behind the scenes to do so, which is far more reliable for calculations than "mental arithmetic".

• Describe the data: what each column means, units, the time period.
• Ask it to show its working — the code or formula used — so you can check.
• Ask it to check data quality first: "List missing values, duplicates or anything unusual before analysing."
• Verify key numbers yourself (see the Validating AI Output course).
• Remember data protection: remove personal data you don't need (AI Literacy, Module 4).

4.3 Multimodal prompting: images, screenshots and audio

Multimodal models can work with images, PDFs with charts, screenshots and sometimes audio and video. Good practice:

• Say what the image is and what you want: "This is a photo of our whiteboard from a retro. Transcribe every sticky note into a table grouped by column."
• Point to specifics: "Focus on the chart in the bottom-right."
• Ask it to describe before it interprets — this exposes misreadings early.
• Check carefully: small text, handwriting, and exact values on charts are common error points.

4.4 Controlling output format

When output will be pasted somewhere else — a spreadsheet, a system, an email — be precise:

Return a JSON array only, no commentary. Each item:
{"name": string, "category": "BILLING" | "DELIVERY" | "PRODUCT" | "OTHER",
 "urgent": true | false}

Many AI platforms also offer structured outputs features that guarantee valid formats for developers. For everyday use, a clear example of the format is usually enough.

4.5 Generating images

For image generation, describe the subject, setting, style, composition and mood, and iterate. Check images for errors (hands, text, logos), bias and IP concerns, and label AI-generated images where appropriate.

Key takeaways

• Put long documents first, label them, and ask your question last.
• Ask for relevant quotes before the answer.
• For data, describe the columns, ask to see the working and verify key numbers.
• With images: say what it is, point to specifics, describe before interpreting.
• Show the exact output format you need.