How to Read an AI Image Prompt: The Six Parts Behind Every Picture

Most people meet a good AI image before they meet its prompt. Say it shows a lighthouse keeper’s kitchen at dawn: a kettle steaming, a cat asleep on a sea chart, cold window light against a warm stove. You want something like it, so you type “cozy lighthouse kitchen, morning” and get a picture that looks like a coffee mug ad.

The difference between the two images is rarely luck. The first prompt was about sixty words long, and nearly every word had a job. The second had four words and left everything else to the model’s defaults.

This guide shows how to take a finished prompt apart, so the next one you write does more of the work itself.

Why reading comes before writing

Writers have learned this way for a long time. In his autobiography, Benjamin Franklin describes taking notes on essays from The Spectator, putting the notes away for a few days, rewriting the essays from them, and then comparing his version with the original. The comparison showed him his faults.

Prompting rewards the same habit. A prompt on its own is a list of intentions. A prompt next to its result shows which intentions the model followed, which it ignored, and what it added on its own. That is why the AIContestList Prompt Library shows every prompt beside the image it produced, and lists the model and version only when they were recorded at generation, never guessed from the picture.

The six parts of an image prompt

Part	What it decides	Words that usually do the job
Subject	Who or what the picture is about	“an elderly lighthouse keeper’s kitchen”, “a ginger cat”
Setting	Where and when	“at first light”, “on a rocky island in winter”
Composition	Where the viewer stands and what fits in the frame	“seen from the doorway at eye level”, “close-up”, “wide shot”
Light	Direction, color and quality of the light	“cold blue window light”, “warm glow from the stove”
Style or medium	What the image looks made of	“gouache with visible brush texture”, “35mm film photograph”
Settings	Shape, size and other controls outside the text	aspect ratio such as 4:5, image size, model, reference image

Not every prompt needs all six. A product shot may skip the setting; an abstract piece may skip the subject. When a part is missing, the model fills it with its own default, and defaults look alike across thousands of images. Much of what people call the “AI look” is a stack of defaults.

A worked example

Here is the full prompt behind the kitchen picture:

An elderly lighthouse keeper's kitchen at first light, seen from the doorway at eye level. A cast-iron kettle steams on a small stove; a ginger cat sleeps on a nautical chart spread across the table. Cold blue light from a salt-streaked window meets the orange glow of the stove. Painted in gouache with visible brush texture and a muted palette.

Settings: aspect ratio 4:5.

Now sort the words:

• Subject: “an elderly lighthouse keeper’s kitchen”, the kettle, the ginger cat, the nautical chart.
• Setting: “at first light” and “salt-streaked window”. The second phrase puts the sea outside without ever showing it.
• Composition: “seen from the doorway at eye level”. One phrase, and the viewer is standing in the room.
• Light: “cold blue light … meets the orange glow”. Two light sources in opposite colors give the picture its contrast.
• Style: “gouache with visible brush texture and a muted palette”.
• Settings: the 4:5 frame, chosen outside the text.

Notice what the prompt leaves out. It never says “cozy”. It describes the things that make a room feel cozy, such as steam, a sleeping cat and warm light against cold, and lets the viewer arrive at the word. Mood adjectives tend to produce generic results. Concrete objects and named light sources produce specific ones.

How to read a prompt in four passes

1. Look at the image for half a minute before reading any text. Write down the first three things you notice.
2. Find the words responsible for each of those three things. If there are none, the model supplied it.
3. Sort the remaining words into the six parts. Words that fit nowhere are often doing nothing.
4. Check the generation details: model, version, aspect ratio, parameters and any reference image. On a Prompt Library page these sit under Generation details and Parameters.

Step two is where most of the learning happens. A detail with no words behind it tells you about the model, not the prompt, and it may not appear the next time you run the same text. Same Prompt, Different Image explains why.

Sentences or keyword lists?

Older prompts often read like tag lists: “lighthouse, kitchen, cat, dawn, gouache, trending, 8k”. Current image models read language closely. Google’s own guide to prompting its Gemini image model puts this first: describe the scene in sentences instead of listing keywords. The examples in OpenAI’s image generation documentation are written the same way, as plain descriptions.

Two habits from the tag-list era are worth dropping. Size words such as “8K” do not set resolution in tools where size is a separate setting; in AI Studio you choose 1K, 2K or 4K outside the prompt. And it usually works better to describe what you want than what you don’t: “an empty street at dawn” gives the model a picture to make, while “no cars” still puts cars in the sentence.

Practice by changing one part at a time

Reading teaches you to see the parts. Changing them teaches you what each one does.

Pick a prompt from the library and open it with Try in AI Studio. Change only the light, from “first light” to “a single bare bulb at midnight”, and run it. Then put the light back and change only the composition, from “from the doorway” to “from directly above the table”. Keep a one-line note of each change and what it did. After ten runs you will have a reference built from your own eyes, which is worth more than any list of “best prompts”.

For this kind of practice a fast, low-cost model at 1K is enough; keep the larger sizes for the version you plan to keep. Nano Banana or GPT Image? covers which model suits which job. If you run out of things to change, Inspiration Lab will deal you a subject, a style and a light at random.

Where to find prompts worth reading

The Prompt Library can be browsed by model, style, use case and color palette, so you can study one variable across many images: twenty prompts in a single style, or every prompt whose result is mostly blue. Each prompt page shows the prompt, any negative prompt, the generation details and parameters, and who made the image and under what license, as far as that is known.

If you plan to reuse someone else’s prompt in published or competition work, read Can You Reuse Someone Else’s AI Prompt? first. For a tour of how the library, AI Studio and the competition listings fit together, start with AIContestList, Explained.

Questions people ask

How long should an AI image prompt be?

Long enough to cover the parts you care about. For models that read full sentences, one paragraph of 40 to 90 words covers subject, setting, composition, light and style. AI Studio’s Enhance tool, for example, rewrites a prompt into one paragraph of at most 90 words; the prompt field accepts up to 2,000 characters.

Do words like “8K” or “masterpiece” improve an image?

Rarely in the way people expect. Where image size is a setting, as in AI Studio, a word in the prompt does not change the pixel count. Describe the detail you want to see instead, such as “visible brush texture” or “fine wood grain”.

What is a negative prompt?

A separate field listing things to leave out, used by some tools such as Stable Diffusion. Where a model has no negative field, describe the scene you want in positive terms instead.

Can I copy a prompt from the Prompt Library?

Yes. Each prompt page has a Copy prompt button, and Try in AI Studio opens the prompt ready to adapt. Expect a different image: results vary between models, versions and runs, even with the same text and settings.

Sources