Prompts for AI images: template, guidelines and examples

A good prompt for AI images describes three things: what is seen (the subject), where and when (the context) and how it looks (the style), with concrete details rather than generic words. Google's guidelines for generating images with Gemini add more: explain what the image is for, describe what you want rather than what you do not want, use camera terms and ask for the image explicitly. Below you will find a template and examples ready to adapt.

The parts of an image prompt

A prompt is the text description of what you want to generate. Google's guide for its Imagen model proposes a simple starting point: think about the subject, the context and the style. And it warns that requests do not have to be long or complex, but most good requests are descriptive and clear.

• Subject: the object, person, animal or landscape you want to appear. The more specific, the better: not “a boat”, but “a wooden fishing boat with peeling blue paint”.
• Context and background: where it is and what surrounds it. The same guide reminds you that the background is just as important as the subject.
• Style: it can be general (painting, photography, sketches) or very specific (pastel painting, charcoal drawing, isometric 3D).

To those three parts it is worth adding two that decide a lot of the result: the light (dawn, lamplight, overcast sky) and the framing (wide shot, close-up, top-down view).

Google's guidelines for Gemini

If you are going to generate images with Gemini, Google Cloud's documentation lists these best practices, with its own examples:

1. Be specific. The more detail you provide, the more control you have. Instead of “fantasy armor”, Google suggests something like “ornate elven plate armor, etched with silver leaf patterns”.
2. Give context and explain your intent. Saying what the image is for helps the model: “Create a logo for a high-end, minimalist skincare brand” works better than “Create a logo”.
3. Iterate. Do not expect to get a perfect image on the first attempt: use follow-up instructions for small changes, such as “Make the lighting warmer”.
4. Go step by step on complex scenes. First the background, then what goes in the foreground, finally the detail.
5. Describe what you want, not what you do not want. Instead of “no cars”, describe the scene positively: “an empty, deserted street with no signs of traffic”.
6. Control the camera. Use photographic and cinematic terms such as “wide-angle shot”, “macro shot” or “low-angle perspective”.
7. Ask for the image explicitly. Start with “create an image of” or “generate an image of”: otherwise, the multimodal model might reply with text instead of the image.

A template for writing prompts

With those guidelines, a template that works for most of the images in a video is this:

Create an image of a [framing] of [specific subject, with details],
[what it is doing or what state it is in], in [place and period],
with [light], [style], [colours].

Filled in for a scene in a history video, it would read: “Create an image of a wide shot of a medieval market square with cloth stalls and fruit baskets, first thing in the morning, with low mist and golden light, watercolour illustration, ochre and green tones”.

Two practical points. If the image is for a video, add the intent (“for a narrated educational video, horizontal format”). And if you are going to make many images for the same piece, keep the style and colours as a fixed sentence that you paste at the end of every prompt: that is what keeps a scene consistent from one image to the next, as we explain in AI images for videos.

Examples of prompts for images

Examples designed for narrated videos of different kinds, with the weak version and the worked version. They are illustrative: adapt them to your style.

What for	Weak prompt	Worked prompt
History	An old castle	Create an image of a stone castle on top of a rocky hill, at sunset, with red banners blown by the wind and a dirt track winding upwards, oil painting, wide shot.
Science	A cell	Create an image of a scientific illustration of an animal cell in cross-section, with a purple nucleus and orange mitochondria, on a dark blue background, textbook plate style, no labels.
Business explainer	An office	Create an image of a small carpentry workshop seen from the doorway, with a tidy workbench and shavings on the floor, natural morning light coming in through a high window, wide-angle photograph, no people.
Trivia	An octopus	Create an image of a macro shot of an octopus camouflaged against a background of rocks and seaweed, its skin imitating their texture, clear water and light coming in from the surface, underwater photograph.
Landscape or travel	A mountain village	Create an image of a village of stone houses with slate roofs on the slope of a valley, in autumn, with mist among the trees, low-angle perspective, soft watercolour.
Abstract concept	The passing of time	Create an image of a glass hourglass on a wooden table, with the sand falling and dry leaves around it, warm side lighting, blurred background, still-life photography style.

Note what the worked ones have in common: a specific subject, a place, a light, a style and a framing. And that none of them asks for text inside the image: in a video, titles and figures are better added in the edit.

Mistakes that spoil a prompt

• Staying generic. “A beautiful picture of nature” leaves every decision to chance.
• Writing in the negative. A list of “without this, without that” works worse than describing the scene as you want it, according to Google's own guideline.
• Contradicting yourself. “Realistic photograph in cartoon style” forces the model to choose for you.
• Cramming too much into one scene. If you need five elements, consider splitting it into two images or building it step by step.
• Asking for real people or celebrities. Besides the problems of imitating someone, Google's safety filters for images detect a photorealistic depiction of a celebrity in the request. If you are wondering what rights you have over what is generated, we cover it in AI images and copyright.
• Not iterating. The first result is rarely the final one: change one thing at a time so you know what worked.

Prompts in the Estudio de Canales

In the Estudio de Canales you do not need to start prompts from scratch: they are part of the video's storyboard.

• The AI proposes the prompt for each scene from its narration. Its instructions ask for a concrete description of what is seen, consistent with the audiovisual style in the channel bible, with no brands, no recognisable real people or celebrities and no logos.
• You edit it before generating the image, with a maximum of 1,000 characters per prompt.
• You do not have to repeat the style. When generating, the studio adds to your prompt the image style from the channel bible and the request for a horizontal 16:9 image with no text, no lettering, no logos and no watermarks.
• Changing the prompt invalidates the previous image: the scene review asks you to generate it again so that what is seen matches what was requested.
• If Google rejects the description, the studio tells you so that you can rephrase it.

What it does not do: it does not edit an image with follow-up instructions (“make the lighting warmer”) or use reference images; each version is generated anew from the prompt and the style. To iterate, you change the prompt and generate that scene again, without touching the others.

Sources consulted

Checked on 11 October 2026. We link to the original source so that you can consult the current text.

• Guía de atributos de imagen y peticiones (Imagen). Google Cloud · documentación.
• Prácticas recomendadas para la generación de imágenes de Gemini. Google Cloud · documentación.
• Generación de imágenes con Gemini y la IA responsable. Google Cloud · documentación.