Nano Banana prompts: change one thing, keep the rest

No sign-up, nothing sent anywhere — the prompt is assembled in your browser. Here is the thinking it is built on.

Why editing prompts are a different craft

Most prompting advice was written for generators — for the box where you type a description and a brand-new image appears. That advice is real, but it quietly assumes you are starting from nothing. Editing inverts the problem. The image already exists; it already has a subject, a light source, a colour grade, a crop, a mood. Your job is no longer to conjure a scene, it is to reach into a finished one and move a single stone without disturbing the rest of the wall. Those are not the same skill, and the prompts that serve them do not look alike.

Here is the trap. Ask a generator for “a woman in a red coat on a rainy street” and any plausible woman, coat, and street is a correct answer — there is no ground truth to violate. Now hand an editor a photo of a specific woman and ask it to “make the coat red.” Suddenly there is a ground truth: her face, her hair, the exact fold of the fabric, the puddle reflecting the streetlight. The under-specified instruction that was harmless for a generator is now dangerous, because the model has to re-draw a region and every unstated detail is up for grabs. It may give you a red coat — and a subtly different person wearing it. You did not ask it to keep her face; so it did not.

That is the whole difference in one sentence: a generation prompt is judged by what it adds, and an edit prompt is judged by what it protects. The best edit instructions spend more words on preservation than on the change itself. It feels redundant — of course you want her face to stay the same — but “of course” is not in the prompt, and the model cannot read the picture in your head of what “done right” looks like. It can only read the words. Say the quiet part.

There is a second shift, subtler and just as important: scope. A generator works on an empty frame, so scope is meaningless. An editor works on a full frame, so scope is everything. “Change the lighting” could mean the whole scene or the one shadow on the left cheek. “Make it look vintage” could restyle the entire image or just the poster on the wall. Every edit needs a boundary — a clear statement of where the change stops. Master those two moves, protection and scope, and you have mastered ninety percent of image editing. The rest is vocabulary.

One more habit separates people who edit well: they change one thing at a time. It is tempting to stack five fixes into a single instruction — new background, brighter subject, remove the sign, warm the colours, straighten the horizon — and let the model sort it out. It almost never does. The requests interfere, and worse, you cannot tell which word caused which regression, so you are left rewriting the whole thing and hoping. Editing is iterative by nature: run one change, look at the result, feed that result back in for the next. A chain of four small, verifiable edits beats one heroic paragraph every time — and when something breaks, you know precisely which step to blame.

The anatomy of a clean edit instruction

A prompt that only changes one thing is not a lucky sentence; it is a small structure with four moving parts. Say all four and the edit lands. Skip one and you have handed the model a decision you did not mean to delegate. The builder above is these four parts turned into fields — but once you can see the shape, you can write it freehand in any box.

1 — Target: name the thing, not the vibe

The target is what in the image is changing, and it must be specific enough that a stranger could point to it. “the object” or “the background” is often fine, but “the chipped blue mug on the left edge of the desk” is safer, because it removes every competing interpretation. When two things in the frame could match your description, the model picks one — and it will not be the one you meant half the time. A precise target is also how you set scope: naming one mug tells the model the other three objects are off-limits without you having to say so.

2 — Change: describe the destination, concretely

The change is the single difference you want, and vague adjectives are where quality leaks. “make it nicer” hands the model your taste to guess at. “make it a matte black ceramic mug with a subtle speckled glaze” gives it a destination it can actually hit. Describe the end state, not the verb — what the pixels should be when it is done, including material, colour, and finish where they matter. And keep it to one change. If you find yourself writing “and also,” you have two edits, and two edits in one breath is how a clean retouch turns into a redraw.

3 — Preserve: the clause amateurs skip

This is the load-bearing wall of every edit prompt, and it is the one people leave out. Preservation is the explicit list of what must survive the edit untouched — identity and face, pose and proportions, lighting and shadow, composition, palette, and everything outside the edited area. You will feel silly writing it, because it states the obvious. Do it anyway. “Keep her face, hair, and expression identical; this must read as the same person” is the difference between a portrait retouch and an accidental deepfake of a stranger. If there is one habit to steal from this whole page, it is ending every edit prompt with a preservation clause.

4 — Constraints: fence the frame

Constraints are the global rules that stop the model from “improving” things you never mentioned. Lock the frame — “keep the same aspect ratio, resolution, and crop” — so it does not silently zoom or re-compose. Add the negative space: “do not add text, watermarks, or new objects, and do not recolour the rest of the image.” Constraints are cheap insurance against the model’s eagerness to help. Together with a preservation clause they form a fence: the change happens inside it, and nothing leaks out. The builder writes this fence for you on every prompt, because it is the part you are most likely to forget and most likely to regret forgetting.

Two things make this structure hold up in practice. First, the order is a priority list, not a formality — a precise target and an honest preservation clause do more for an edit than the most poetic description of the change. Second, the four parts are not all sentences; a tight target can carry the scope on its own, and a single preservation line can stand in for a paragraph of hedging. You will not always write four separate blocks. You just have to make sure none of the four is silently left to the model’s discretion, because the part you leave blank is exactly the part that comes back to surprise you.

Five edit recipes you can steal today

Structure tells you why; recipes tell you what to type. These five cover the overwhelming majority of real editing work, and each is just target-change-preserve-constraints applied to a common job. Steal them, fill in the brackets, and keep the ones that earn a place in your library.

Change the background to [new scene]. Keep the foreground subject exactly as-is — same pose, edges, and proportions — and relight it to match the new scene’s light direction and colour temperature, including a grounded contact shadow. Do not alter the subject.

Remove [object] from the image. Rebuild what would be behind it by continuing the surrounding [surface / texture / pattern] seamlessly — no blur, ghosting, warping, or leftover outline. Leave everything else untouched.

Restyle the image as [target style]. Keep the composition, the subject’s proportions, and the position of every element recognizable — change the rendering, not the scene. Same crop and framing.

Clean up [product]: [specific fix — remove dust, even the lighting, straighten the label]. Preserve the product’s true colour, material, shape, and label text exactly. Make no other changes to the shot.

Using the person in this image, [new action or scene]. Keep their face, hair, skin tone, and body proportions identical to the source — this must read as the same individual, changing only [what changed].

Two rules keep these recipes reliable once they are in your hands. First, resist merging them: a background swap and a product retouch are two recipes because they are two edits, so run them in sequence on the same image rather than as one tangled request. Second, treat the brackets as the only thing you change. The value of a recipe is the language around the brackets — the relight instruction, the reconstruction clause, the frame lock — which is the part that took someone a dozen failed tries to word correctly. Swap the target and the change; leave the scaffolding alone. That is exactly what forking a prompt preserves, and why a good library beats a good memory.

When the edit goes wrong: read the symptom

A bad edit is almost never random — it is a symptom, and every symptom points straight back to a missing part of the anatomy. The people who get clean edits do not conclude the model is dumb when it slips; they treat the output like a bug report, change one thing in the prompt, and run it again. Here is the lookup table for the failures you will actually hit.

Notice the pattern: every fix is a single, named clause, because every symptom is a single, named gap. That is the quiet superpower of a structure — it turns “this looks wrong” into “the preservation clause is missing,” which you can actually do something about. Rebuild the prompt in the tool above with the missing part added, and the edit that fought you twice usually lands on the third try. And once it lands, you never have to solve it again — because you can save it.

There is a compounding payoff hiding in that table. Each time you diagnose a symptom and add the clause that fixes it, you are not just repairing today’s image — you are hardening a prompt you can reuse. The background swap that finally stopped haloing is now the background swap, full stop. Save it, and the next one starts from a version that has already survived its mistakes. Over a few weeks, the edits that gave you the most trouble quietly become the most reliable prompts in your library, and the failures above stop being failures you repeat.

From one good edit to a system

Here is the part nobody tells you: writing one clean edit prompt is a skill, but never writing the preservation clause twice is a superpower. The sentence that describes your change is different every time — a mug here, a sky there. But the scaffolding around it — the preserve list, the frame lock, the negative guardrail — is nearly identical across a hundred different edits. The people who ship the most images are not retyping that fence every morning. They keep a library, reach for the closest proven edit, and change only the specifics. That is the entire idea behind PromptFork, and it maps onto three moves.

And when the closest prompt still is not close enough — a strange target, an edit you have never done before — that is what Studio is for. You describe the edit in plain language and it runs the same target-change-preserve thinking you just read about, handing back a precision instruction you can use, refine, and publish back to the library. It is the “Supercharge with AI” button on the builder above: your rough edit in, a forged one out, five free every day. The builder teaches you the moves; Studio does them at speed; the library means you only ever solve each edit once.

And it compounds. In week one, your library is a handful of forked edits. A month in, it is the scaffolding for most of what you produce — the product shots that stay accurate, the portraits that keep the same face across a series, the backgrounds you swap without a telltale halo. You stop starting from a blank box and start from the best version you have found so far, then improve it. That, in the end, is the whole difference between people who “use an image editor” and people who quietly get twice as much out of it: not a better model, but a better starting line.

That library does not have to be elaborate to earn its keep. Start it with the five recipes above and the one edit you do most — each saved the moment it works, each labelled with the job it solves. The point of keeping it on PromptFork is that you are not the only one filling it: you fork from people who cracked a swap or a retouch before you, and the edits you refine become the ones someone else starts from next month. A prompt library is the rare asset that grows more useful the more it is shared, because every fork is a quiet field test. The blank box is where everyone begins; a reusable library is how you stop beginning there.

Nano Banana prompts worth forking right now

Theory is cheap. Here are real, community-tested prompts you can copy or fork this minute — each one is a clean edit instruction in the wild.

Questions people ask about Nano Banana prompts