TL;DR
A vision model doesn't "get" your photo the way a person does — it reads it as a grid of visual evidence, so your prompt must name the subject, the action, and any visible text to aim its attention. And because decoding an image burns vastly more computing than decoding words, every attachment carries a real cost you should spend only when the image itself is the data.
The Analogy
Picture a brilliant scholar who has spent their entire life blindfolded.
They have read every book ever written — medicine, law, art history, chemistry — entirely through touch and text. Then one day, the blindfold comes off. The eyes work. But here's what didn't happen: no one handed them a lifetime of visual experience. They have no instinct for what matters in a scene, no reflex that says "everyone looks at the face first." Open-eyed, they are a genius with no idea where to look.
So when you show them a photo of your kitchen and say "thoughts?" — they panic-study everything. The faucet. The window. A smudge on the third tile. The scholar has the knowledge to interpret any of it, but with no human's sense of the salient, they distribute genius evenly across the whole scene. The answer you get back is exhaustively observant and completely unfocused.
Your words are the finger pointing at what matters. "The sticker on the bottom of the pan — what does it say about the coating?" — and suddenly the world's most educated eyes lock onto exactly the thing you needed.
One more detail, and it explains the price tag: keeping that blindfold off is expensive. Reading an image takes the scholar a hundred times the effort of reading a sentence. So every minute of open-eyed staring is billed at a premium — and you pay it whether the photo was evidence or just decoration.
How It Works & The Cost (Uzu)
When you attach an image, the model slices it into a grid of patches and converts each patch into the same currency it uses for words: tokens. A photo can easily cost hundreds of tokens before your question is even read — and like everything in the conversation, it stays on the context window, the desk the model re-reads before every single reply. A picture isn't just worth a thousand words; on that desk, it occupies them.
That's why visual prompting has its own anatomy — three slots that aim the model's attention:
| Slot | The question it answers | Example |
|---|---|---|
| Subject | What thing in the image is this about? | "the warning label on the charger" |
| Action | What is happening, or what should be done with it? | "translate it" / "is it damaged?" |
| Text | Which visible words matter? | "read the serial number starting with SN" |
Name the subject and you stop the model from averaging its attention over the whole grid. Give the action — translate, identify, compare, estimate — and it becomes a one-shot prompt with eyes. Point at visible text and you exploit a vision model's single sharpest skill: modern models read text inside images better than almost anything else they do.
Then there's the bill. Reading those hundreds of image-tokens takes real GPU time, which is why Uzu charges a flat 10 G-Credits per attached file, priced at the moment you hit send. It isn't a fee for storage — it's the cost of the electricity and compute the model spends staring. The design lesson baked into that price: attach the image only when the image is the data. A photo of the error dialog is data. A screenshot of a paragraph you could have pasted as text is decoration with a tax — the same answer, delivered for the price of plain words. And per limiting scope, one tight visual question per attachment beats five vague ones: constrain the subject, the action, and the output format before you send.
The working rule:
A picture is worth a thousand tokens. Make sure the model spends them where you'd spend your own attention — and only pay for pixels that carry information.
Before & After (The Prompts)
Example 1 — the mystery object
[photo of a cluttered drawer] what's this?
The model dutifully inventories the drawer: cables, coins, a battery, "possibly a small remote." Somewhere in that list is the thing you meant. Probably.
[same photo] Identify the small black device in the upper left, next to the batteries. Tell me its likely brand and what it's for, in 2 sentences.
Subject (the black device, upper left), action (identify brand and purpose), format (2 sentences). Same photo, surgical answer.
Example 2 — the document
[photo of a form] Can you look at this?
"Looking" is not a task. The model summarizes the paper generically — which is fine, unless you needed the one field that decides everything.
[same photo] Read only the "Effective Date" field in the top-right box of this contract page and transcribe it exactly. If it's illegible, say "illegible" — do not guess a date.
Text-slot prompting at its sharpest: one field, exact transcription, and an explicit no-guessing guard — because a model that fills gaps seamlessly is exactly the wrong tool for un-guessed legal dates.
Example 3 — the decoration trap
[screenshot of an article, attached] Summarize this.
You just paid image prices for text. The model reads pixels that spell words — slower, costlier, and more error-prone than if you'd pasted the words themselves.
[no image] Summarize this article in 3 bullets: [pasted text]
Identical answer, a fraction of the compute, zero attachment charge. The screenshot was never data — it was a container for data. This one habit — paste when it's text, attach when it's pixels — is the single biggest saving in visual prompting.
Common Pitfalls
- Expecting the AI to "feel" the image instead of reading it. A vision model doesn't experience your sunset photo; it catalogs evidence from a grid. Ask "what emotions does this evoke?" and you'll get a confident description of conventions — warm tones, smiling faces — not a felt reaction. Ask what's in the image, not what it's like to see it.
- Attaching decoration. Screenshots of text, logos "for context," photos that repeat what you already typed. Every attachment is premium compute spent twice if the information was already in words. Attach only what only the image can carry.
- Vague deixis — "this," "it," "that part." In text chat, "it" points at the last message. In an image, "it" points at everything. Use nouns and positions: the dial on the left, the second line of the receipt, the red wire at the bottom.
- Skipping the no-guessing guard on critical reads. Serial numbers, dates, doses, prices: a vision model under uncertainty fills gaps as smoothly as any other model. Add "transcribe exactly; if unclear, say 'illegible'" — the same honesty-renting trick from the hallucination lesson, now with eyes.
- One image, five questions, one message. Bundling "what's this, what's it worth, where's it from, is it real, write a listing" dilutes attention across the whole grid five times over. One attachment, one sharp question — then follow up; the image stays on the desk, and the follow-ups ride on tokens you already paid for.
FAQ
How do I prompt AI with an image?
Name the subject, the action, and any visible text, then constrain the output: "Identify [the specific thing] in [the location in the image], [do this with it], answer in [format]." Point the model's attention like you'd point a colleague's eyes — at the sticker, not the kitchen.
Why does sending an image to AI cost more?
Because the model must convert the image into hundreds of tokens of visual evidence before it can process your question — dramatically more compute than the equivalent sentence. In Uzu, that premium is a flat 10 G-Credits per file: the honest price of the GPU time spent staring. Text is cheap; pixels are processed.
Can AI read text inside images?
Yes — and it's one of vision models' strongest skills, from handwriting to street signs to screenshots of forms. The catch is accuracy under uncertainty: for anything critical (dates, serials, doses), demand exact transcription and an explicit "say illegible if unclear" instead of letting the model fill gaps.
Next Lesson
Notice what ran through this entire lesson: you did the pointing. You chose the subject, you aimed the attention, you supplied the constraints. But there's a move that flips the whole conversation — what if, instead of answering questions, the AI asked you the questions first? One cheap round of the machine interviewing you can replace three expensive rounds of you re-explaining.
Continue to the next lesson: The "Ask Me First" Technique
Practice drill before you go: find the last photo you sent an AI and re-ask the question three ways — vague, medium, and surgical (subject + action + text + format). Watch the answer sharpen with each one, then check what the vague version actually cost you.