Five prompts, one tool, one place to run them. AI Central's Nano Banana carousel gives you five copy-and-paste prompts, an ordinary iPhone selfie, a multi-image fusion, a 1/7 scale action figure, a crocheted chibi doll and a sheet of chibi emoji stickers. The instruction for running them is one line, go to aistudio.google.com and type the prompt in. What makes them work is over-specification, each one describes the artefact rather than the subject.
The setup is one line, and that is the point
Reviewed August 2026. Most prompt guides spend half their length on configuration. This one spends a single slide. Go to aistudio.google.com, and in the document's own instruction, simply type in your prompt.
That brevity is a real editorial decision. Nano Banana is reached through a plain text box in Google AI Studio, so all the leverage sits inside the prompt string. AI Central states the goal on the opening slide.
Using Nano Banana in a smart way it’s our goal
Smart, in practice, means long. The shortest of the five still runs to nearly forty words.
What the five prompts actually cover
The set handles five jobs, running from the most personal to the most reusable.
- An ordinary iPhone selfie, a deliberately bad snapshot of named people at a named location.
- Image fusion, several source images merged into one coherent scene.
- A 3D action figure, a 1/7 scale collectible photographed on a desk beside its packaging.
- A chibi knitted doll, a crocheted version of an uploaded character held in two hands.
- A chibi emoji sticker sheet, a full set of expressions from one photo, framed 9 by 16.
One cataloguing quirk, the document promises five and delivers five, but the third slide is headed only 3D Action Figure, so the numeral three never appears. Harmless in order, awkward if you file prompts by number.
Prompt one works by describing the flaws
The selfie prompt is the most instructive of the five because it asks for failure. Slight motion blur. Uneven light. Mild overexposure. A bad angle. Image models default to good composition, so the only way to get a snapshot is to name every way a snapshot is not a photograph.
The angle is awkward, the composition is messy, and the overall aesthetic is deliberately plain
That line is from the prompt as AI Central published it. Realism in generated images arrives through named imperfections, never through the word realistic.
It closes with two bracket slots, one for the people in frame and one for the place, plus a fixed detail, taken at night. Night is doing real work there. It forces the flat, harsh, over-lit look the eye reads as an actual phone camera.
Prompt two is a constraint list, not a request
The fusion prompt merges several images into one. Most people would write combine these images and stop. AI Central's version spells out the conditions under which a merge looks convincing.
Blend the images naturally with consistent lighting, shadows, perspective, and style.
It also insists every key subject stays recognisable and that proportions hold. Those two clauses stop a fusion collapsing into an average of faces, the standard failure mode here, and they are the part worth stealing.
Props do the heavy lifting in prompt three
The action figure prompt barely describes a figure. It describes a photograph of one in a real workspace. A computer desk. A circular transparent acrylic base with no text. A BANDAI style packaging box printed with the original artwork. Then one detail almost nobody would add.
On the computer screen, display the ZBrush modelling process of the figure.
A monitor showing the sculpt in progress is the tell that makes the scene read as a photo from a modelmaker's desk. Set dressing sells the render.
The document reports running this one on a Mr.Beast image and calls the result amazing. That is the only outcome claim in the carousel. No comparisons, no failure examples, no attempt count, so treat it as a practitioner's report rather than evidence.
Prompts four and five are the ones you will reuse
The knitted doll prompt turns an uploaded character into a handmade object held by two hands. Its weight sits in texture and light, not in the doll.
The background is slightly blurred, depicting an indoor environment with a warm wooden tabletop and natural light streaming in from a window, creating a comfortable and intimate atmosphere.
Notice what gets specified. The room, the tabletop, the window light, and elsewhere the finger positions and skin texture of the hands doing the holding. The character itself gets one clause. AI Central pitches this as art in a few clicks, which undersells a precise piece of scene construction.
The sticker prompt is different in kind. It is a batch instruction. One run produces a peace sign with a wink, a tearful face, an open armed hug, a sleeping pose on a tiny pillow, a confident point, and a blown kiss trailing hearts. Then it locks the style across all of them.
Maintain the chibi aesthetic. Exaggerated, expressive big eyes. Soft facial lines.
It finishes on production detail, a vibrant red background with stars or confetti, clean white space around each sticker so the sheet can be cut apart, and an aspect ratio of 9 by 16. It is the only one of the five returning an asset pack rather than a single picture.
Why the pattern holds
Read the five side by side and the same method shows up in each.
- They describe the output object, a snapshot, a collectible, a doll, a sticker sheet, not the person inside it.
- They name the lighting and the surface every single time, because lighting is the first thing a model gets wrong.
- They carry bracket slots for names, locations, image references or an uploaded character, so each prompt is a template you fill, not a sentence you rewrite.
- They end on a deliverable spec, a base with no text, packaging beside the figure, white space around each sticker, a fixed aspect ratio.
That last habit separates a picture from something you can ship. A prompt ending in a specification returns a file you can use. A prompt ending in a vibe returns something you redo.
What the document leaves out
Worth being straight about the gaps. The carousel gives no model settings, no guidance on what to change when an output misses, and no examples of a prompt failing. It is a starter set and presents itself as one.
The closing slide is honest about that, and carries the sharpest line in the document.
Nano Banana is a tool.
The line straight after it is the argument for treating prompts as durable assets rather than one-off inputs.
These Prompts, your engine.
What to do with this
If you have never generated an image, run the selfie prompt unchanged with two real names and a real place. The gap between what you asked for and what came back is the lesson.
If you already generate images, strip the sticker prompt for parts. Its structure, a pose list plus a style lock plus a background rule plus an aspect ratio, transfers to product shots, avatars and icon sets once you swap the nouns.
If you are building a repeatable workflow, treat all five as templates, not examples. Save them with the brackets intact, record which slot expects a name and which expects an uploaded image, and version them when you change a clause. AI Central's order of operations is blunt, try the prompts, polish the output, get your best content.
What is Nano Banana and where do I run it
The document treats Nano Banana as an image generation tool and names one place to use it, aistudio.google.com, which is Google AI Studio. You type the prompt in and run it. No other setup step is given.
Do I need to edit the prompts before using them
Only where there are brackets. The selfie needs names and a location, the fusion needs image references, the knitted doll needs an uploaded character. Everything else stays as written, because those fixed clauses produce the look.
Will these work with any photo
The document makes that claim for one of the five, the chibi emoji sticker sheet, which it says is available for any of your photos. It does not extend it to the other four, so test rather than assume.
Why are the prompts so long
Because the length is the technique. Each prompt front-loads decisions the model would otherwise make for itself, lighting, angle, texture, background and format. Shortening them hands those decisions back, which is how you get generic output.