Use AI better
How to use photos, files, and voice with AI
Stop typing everything manually. Give the assistant the material you already have when the product supports it.
The short version
The easiest AI workflow is often to provide the original material — a photo, document, spreadsheet, screenshot, or spoken explanation — and then ask for one useful transformation.
Optional explanationRead the full guide
Photos are useful when the visual detail matters
A photo can save you from describing every object, label, ingredient, or visible symptom. It is especially useful for everyday inspection and idea-generation tasks.
- Photograph a fridge or pantry for meal ideas.
- Show a model label plus the problem when troubleshooting a household item.
- Share a room or object when you need help describing, organizing, or identifying what to inspect.
Files are useful when accuracy depends on the source material
If you want a document summarized or compared, giving the AI the actual document is usually better than describing it from memory.
- Compare contractor estimates side by side.
- Turn several routine messages into one checklist.
- Pull decisions and actions from notes or a transcript.
Good instruction
Use only these files. Separate what the documents explicitly say from your interpretation. If something is missing, label it “not stated” instead of guessing.
Voice is useful when talking is easier than typing
You can often get better raw material by speaking naturally for a minute than by staring at an empty text box. Talk through the situation, then ask the AI to organize what you said.
- Dictate rough notes for an email.
- Talk through a project idea and ask for the first five steps.
- Explain what happened with a household problem in chronological order.
Always inspect what you are sharing
Photos and files often contain more information than you intended to share: names, addresses, account numbers, children’s information, metadata, background documents, or other private details. Crop, redact, or remove unnecessary material first.
Give the material one clear job
Uploading a file and saying “analyze this” forces the AI to guess what matters. Tell it the transformation you want: compare, summarize, extract dates, find differences, turn into questions, or explain a specific section.
Better than “analyze this”
“Use these two estimates only. Put them into the same categories, mark anything not stated, and give me the five differences I should ask each contractor about.”
Photos can be input, not just decoration
One of the most useful mindset shifts is realizing that you do not always need to describe the situation in words first. A photo of ingredients, a model label, a receipt, a whiteboard, or handwritten notes can become the starting material for a real task.
The photo still needs context. Tell the AI what you care about and what it should not infer from the image alone.
Voice can turn thinking out loud into something usable
If you are stuck because writing feels like work, talk for a minute instead. Explain the situation naturally, including the annoying details you would probably omit from a formal prompt, then ask the AI to organize the result.
- Turn a rambling project explanation into the first five actions.
- Talk through what happened in a meeting and ask for a clean follow-up draft.
- Describe what you need from a difficult email before asking for the actual message.
Optional depthGo deeper when you want to
Multimodal input is powerful because it can remove description work. The deeper skill is preserving source fidelity: making sure the assistant knows what came from the material, what you supplied as context, and what it inferred.
Define the source boundary before asking for interpretation
When accuracy matters, tell the assistant which material is authoritative and how to represent missing information. This prevents an apparently helpful model from filling gaps in a way that looks like extraction.
- Use only the supplied file/image/transcript for factual claims.
- Mark missing fields as ‘not stated’ rather than guessing.
- Separate visible/extracted facts from interpretation.
- When comparing multiple files, identify which source supports each important difference.
Source-fidelity instruction
‘Create a table from these two quotes. Every row must say included, excluded, not stated, or unclear, and identify the quote/page where you found it.’
Use stable output schemas for repeated document work
If you repeat the same document task, do not only reuse the prompt. Reuse the output structure. Stable columns and headings make omissions visible and make different runs comparable.
That is especially useful for estimates, bills, receipts, transcripts, applications, notices, and other semi-structured material where the important fields recur.
- Define required output sections.
- Define explicit missing/uncertain states.
- Require source location for consequential fields.
- Keep interpretation in a separate section from extraction.
Treat image understanding differently when the consequence rises
A photo is excellent for orientation, identification clues, and deciding what to inspect next. It is much weaker as sole proof for safety, diagnosis, authenticity, or a condition that requires measurements or hidden information.
The deeper habit is to ask what the image cannot show. Then use the answer to decide whether you need a label, measurement, another angle, official manual, sensor, test, or qualified person.
Better visual reasoning
Instead of ‘Is this safe?’, ask ‘What visible clues can you identify, what cannot be determined from this photo, and what should I verify before deciding whether it is safe?’