How I Reverse-Engineer Any AI Image Into a Prompt I Can Actually Reuse

You've seen the image. The one that stops your scroll. A product shot lit like a film still, or a portrait with color so specific it feels like a memory. You want to make something in that style. So you describe it to your model in plain English — "a moody cinematic photo of a watch" — and what comes back is fine. Generic. Nothing like the thing that stopped you.

The problem isn't your model. It's that you described the *vibe* and skipped the *recipe*.

A great image isn't one idea. It's six or seven decisions stacked on top of each other, and most of them are invisible until you go looking. When I find an image I want to learn from, I don't try to copy it. I take it apart. I figure out which decisions made it work, write each one down as a piece of a prompt, and rebuild it as something I can run on my own subjects. The image is disposable. The recipe is the asset.

Here's how I do it, and the mental model that makes it fast.

An image prompt has anatomy

This is the whole trick. Once you see that every strong image is made of the same parts, you stop staring at the picture as a single magic object and start reading it as a list of choices.

The six parts I look for, every time:

Subject. What's literally in the frame. The watch, the woman, the empty diner.

Style. The aesthetic lineage. Film photography, 3D render, oil painting, anime, brutalist poster. This is the part people skip, and it's the part doing the most work.

Lighting. Where the light comes from and how hard it is. Soft window light, harsh direct flash, golden hour, neon spill. Lighting is what separates "a photo of a face" from "a portrait."

Composition. How the frame is arranged. Close-up, wide shot, the angle, where the subject sits, how much empty space surrounds it.

Medium and render. What it's pretending to be made of. Shot on 35mm film, rendered in Octane, hand-drawn ink, claymation. This is texture and grain and the fingerprint of a tool.

Parameters. The technical knobs. Aspect ratio, a lens length, a color grade, a stylization setting.

You don't always get all six from a single image. But every time you name one more of them, the gap between "generic" and "the thing that stopped you" gets smaller. Most failed prompts are just a subject with no style and no light. You wrote the noun and forgot the other five.

Reading an image part by part

Let me show you the actual move. I'll describe an image, then walk through what I pull out of it.

Picture a coffee cup on a windowsill. But it's not a stock photo. The light is low and warm, raking across the table from the side. There's grain in the shadows. The colors are slightly faded, like an old photograph. The cup is off to one side with a lot of empty, soft-focus room behind it.

Plain-English me would type "cozy coffee cup by a window." Here's what anatomy-me writes instead:

A ceramic coffee cup on a wooden windowsill, steam rising. Warm low-angle morning light raking from the left, long soft shadows. Shot on 35mm film, fine grain, slightly faded muted color palette, gentle vignette. Off-center composition, shallow depth of field, soft-focus background with negative space on the right. Aspect ratio 3:2, shot at 50mm.

Same cup. But now I've named the light (warm, low, from the left), the medium (35mm film, grain), the color treatment (faded, muted), the composition (off-center, shallow, negative space), and the parameters (3:2, 50mm). The "cozy" feeling I started with wasn't one thing. It was five things, and now each one is written down where I can change it.

That's the unlock that isn't an unlock. The feeling becomes a list. Lists are reusable. Feelings aren't.

Rebuild it on your own subject

Here's why this matters more than just copying one picture. Once you've broken an image into parts, the subject is the only piece that's truly stuck to that original. Everything else — the style, the light, the medium — is a style engine you can point at anything.

So I do exactly that. I keep the recipe and swap the subject.

Take a portrait I admired: a person photographed against a deep shadow, lit by a single hard light from one side, half the face in dark, rich skin tones, the texture of real film. I pulled the recipe out of it. Then I pointed it at a product instead of a face:

A matte black mechanical keyboard on a dark surface. Single hard key light from the upper left, deep shadows, dramatic chiaroscuro, half the subject falling into black. Rich contrast, true blacks, fine film grain, shot on medium format. Tight composition, low angle, the object emerging from darkness. Aspect ratio 4:5.

The subject changed from a face to a keyboard. The lighting recipe — one hard light, deep shadow, chiaroscuro — came along untouched, because that was never about the face. It was about the *light*. That's the part worth keeping. The keyboard shot looks like it belongs in the same world as the portrait, and I never had to recreate the portrait at all.

This is the difference between stealing an image and learning a style. One gives you a copy. The other gives you a tool you can run on your whole catalog.

A third pass, for the harder case

Styles get easier to read with reps, but the abstract ones fight back. Say you find an illustration: a retro travel poster, flat shapes, a limited palette of three or four colors, thick clean outlines, a grainy printed texture like it came off an old press. No photography words apply here. There's no lens, no film stock. So I read the medium differently.

A retro travel poster of a mountain range at sunset. Flat vector illustration, bold geometric shapes, limited palette of four colors — burnt orange, cream, deep teal, navy. Thick clean outlines, screen-printed texture with visible grain, slight off-register color. Minimalist composition, strong horizon line, lots of negative sky. Vintage mid-century poster style, portrait orientation.

The anatomy still holds. Subject (mountains at sunset), style (mid-century poster, flat vector), "lighting" (sunset, which here is really a color story), composition (strong horizon, negative space), medium (screen-print, grain, off-register), parameters (four-color palette, portrait). The labels bend to fit the kind of image, but the skeleton is the same six questions every time. That's what makes it a system instead of a guess.

Where I cheat to go faster

I'll be honest about the slow part. Reading all six parts off an image by eye takes practice, and on a busy day I don't want to sit there squinting at someone's lighting setup. This is the exact thing I built into Super Prompts, because I was doing it manually a dozen times a week.

I drop the image in, and it gives me back the prompt broken into its parts — subject, style, lighting, composition, medium, parameters — already separated and labeled. Not one blob of text. The actual anatomy, sliced. From there I do the real work: keep the recipe, swap my subject, save the version that worked. If you want the deeper version of that last step, I wrote about turning a good prompt into a reusable template — the image case is the same move, just with pictures as the starting point.

The point isn't that a tool reads the image for you. It's that once it's in parts, you can *think* in parts. And thinking in parts is the whole skill.

What to try this week

Find one image you wish you'd made. Don't describe it as a feeling. Open a note and write six lines: subject, style, lighting, composition, medium, parameters. Force yourself to fill in all six, even when you're guessing. The guessing is where you learn what you've been leaving out.

Then change line one. Keep the other five, point them at your own subject, and run it.

You'll notice the gap close. Not because you found a magic phrase, but because you finally wrote down the five decisions you used to skip. A stunning image you saved to a folder is inspiration you'll forget by Friday. A stunning image you took apart is a recipe you'll cook for the rest of the year.

When you find a breakdown that works, save it where you can run it again — Super Prompts is free to start. The best style you reverse-engineer is worthless if you can't find it next month.