How to write a good prompt: 12 rules for images

Short answer: a good image prompt describes the scene not as a pile of keywords but in a few sentences built from concrete, showable details: who, where, doing what, where the light comes from, where the camera is. Google's Gemini documentation and the Black Forest Labs (FLUX) guide point in the same direction. Below are 12 rules, each with a before-and-after example; they are not tied to one tool and work in most models.

What should an image prompt contain?

BFL's guide suggests this skeleton to start with: subject, location, style, camera settings, lighting, colors, effect and additional elements. It adds that this is a starting point, not a strict formula. The guide says short prompts work too, and that you should add detail only where you want control. The Gemini documentation agrees: the more richly you describe the scene, the more control you have over the result.

The 12 rules

1. Describe a scene, not a keyword list

The core recommendation in the Gemini documentation is to describe the scene in detail and in flowing language. Instead of "cat, window, sun", build the relationships.

Before: cat, window, sun, cozy

After: A gray cat sleeps on a wooden windowsill while afternoon sun falls across its back and the curtains.

2. Name the subject first

The BFL guide says to start with the subject and then add the details you want to control. When the subject opens the sentence, the model more easily understands what is central.

Before: At sunset near the sea, there is an old fisherman with a boat and nets.

After: An old fisherman mending a net beside his wooden boat, on a rocky beach at sunset.

3. Write showable details instead of "beautiful" and "masterpiece"

BFL recommends writing details you could point to and avoiding vague praise such as "beautiful", "iconic" or "masterpiece". Every phrase should name something that can be shown in the frame.

Before: a beautiful masterpiece of a kitchen

After: A small kitchen with white tiles, a copper kettle on the stove, and a bowl of lemons on a wooden counter.

4. Give context and purpose

Saying what the image is for helps the model choose tone and composition: a warm photo for a café menu is not the same as a tech poster.

Before: a coffee cup

After: A product photo of a ceramic coffee cup for a small café's menu, warm and inviting, space on the left for text.

5. Name the light

BFL recommends writing the direction and quality of the light; harsh, directional light creates long, sharp shadows, while soft light does the opposite. Instead of "good lighting", describe the source.

Before: a portrait with nice lighting

After: A portrait lit by soft window light from the left, gentle shadows on the right side of the face.

6. State the framing and the camera

The photorealistic template in the Gemini documentation specifically lists shot type, camera angle and lens. Phrases such as "wide-angle", "macro" and "low angle" shape the composition.

Before: a mountain road

After: A wide-angle shot from a low angle of a winding mountain road at dawn, mist in the valley below.

7. Say the medium and style explicitly

Photograph, watercolor, vector drawing? In BFL's sample captions, medium and style come first. The Gemini documentation also recommends writing visual detail such as line weight and color for stylized work like stickers or illustrations.

Before: a fox in the forest

After: A flat vector illustration of a red fox in a pine forest, bold outlines, limited palette of orange, teal and cream.

8. Put the text in quotes

The BFL guide recommends putting the exact words in quotation marks and then saying where they go and how they look. The Gemini documentation also asks you to describe the text, font style and color scheme. Keep the text short and check the output letter by letter.

Before: a poster for a book sale

After: A poster with the headline “BOOK SALE” in bold sans-serif capitals at the top, and “Saturday 10-4” in smaller type below, on a cream background.

9. Write what you want, not what you do not

BFL's FLUX 3 Image documentation has no negative prompt field, and when you do not want color it suggests positive terms such as "black and white". The Gemini documentation similarly recommends a positive description such as "an empty street" instead of "no cars".

Before: a beach with no people

After: An empty beach at early morning, smooth untouched sand and gentle waves.

10. Give reference images a role

When you supply several images, BFL recommends referring to each in order ("image 1", "image 2") and stating its role. Without a role, the model guesses which is the character and which is the style.

Before: combine these two pictures

After: Place the person from image 1 into the scene of image 2, keeping the lighting of image 2.

11. When editing, say what stays

In the editing examples in the Gemini documentation, elements that must not change are stated explicitly. A sentence such as "change only the background; keep the face, hair and clothing the same" reduces unwanted drift.

Before: make it sunset

After: Change only the sky to a warm sunset; keep the person, clothing and framing exactly the same.

12. Change one thing at a time and iterate

The loop BFL recommends: start simple, check what is right and wrong, and fix one detail at a time. The Gemini documentation also recommends multi-turn conversation for iterating. Use low resolution for quick drafts and take the final at a larger size.

Before: (five changes at once) warmer, brighter, wider, add a dog, remove the car

After: Round one: "add a dog next to the bench". Round two: "make the light a little warmer". Round three: "remove the parked car on the right".

What if the rules conflict?

Some rules can work against each other: adding detail increases control, but an overlong prompt can confuse the model. BFL's advice is practical: most images do not need a long description; start with the subject and add only the details you want to control. If the result is not what you expected, change a single detail and try again instead of lengthening the prompt.

There is also the question of repeatability: BFL writes that FLUX 3 Image has no seed field and recommends running the same prompt more than once before deciding a change did not work. Do not judge from a single result.

Which language should you write in?

BFL shows that non-English text given inside quotation marks works (the example uses “東京”). The Gemini documentation lists specific languages for best performance, and Turkish is not on that list. The practical approach: write the prompt in English, give any Turkish text that should appear in the image inside quotes exactly, and check the Turkish characters in the output.

Common mistakes

• Piling up filler words: Instead of "8k, masterpiece, ultra detailed", write details that can be shown in the frame.
• Contradictory phrases: "Dark night" and "bright sunlight" in the same prompt pull the model two ways.
• Giving text without quotes: Unquoted text may be read as part of the description.
• Asking for many changes in one prompt: Go in sequence and change one thing per round.
• Not giving references a role: Write what each image is for.
• Deciding from a single result: Try the same prompt a few times.

Summary

• Write sentences that describe a scene, not a keyword list; put the subject first.
• Name the light, framing, medium and style; avoid vague praise like "beautiful".
• Put text in quotes and describe what you want in positive terms.
• Give reference images a role; when editing, say what stays.
• Change one thing at a time, test drafts small, and compare several results.