AI Image & Video

AI Image Prompts: Six Rules Plus Four Questions to Answer Before You Write

When a result is wrong, most people add adjectives, switch to a bigger model, or keep rerolling. All three are expensive and often useless. The real problem is descriptions that conflict, or phrasing that fights how the model works. Fix the seed, change one variable at a time, and you can find which two lines are fighting.

10/17/2026 ｜ 9 min read ｜ Works for Midjourney, Stable Diffusion, ComfyUI and hosted tools. About 8 minutes to read

The one idea that matters

When AI image generation fails, it is usually not the model. Your descriptions are fighting each other.

When a result is off, most people pick one of three moves: add more adjectives, switch to a bigger model, or keep rerolling. All three cost money or time and often do nothing, because the problem is not “not enough description.” It is descriptions that conflict with each other, or phrasing that goes against how the model works.

The way to check this is a controlled experiment: the same model, the same random seed, the same sampling parameters, and only the wording of the prompt changed. Three scenes, two versions each, six images in total. Two clearly got better, one clearly got worse, and the one that got worse produced the most valuable rule.

Controlling variables is the whole point. If you also change the seed, you can never tell whether the prompt improved or you simply got lucky. To test any prompt technique, lock the seed first. Most tools have this option in their advanced settings.

Six rules

Rule 1: In one image, give at most one character a complex pose

The trap: I wanted “two people facing off over the hood of a car, guns raised.” The first version, written simply, worked. For the second I wanted more tension and gave each character a complex pose of their own: one “standing, arms extended over the hood,” the other “back against a wall, half-crouched behind a wheel, gun hand trembling.”

Result: the model blended the two descriptions and reinterpreted them. “Arms extended” was read as a surrender, “half-crouched” was read as crouching to pick something up, and both guns disappeared.

What to do:

• Give the main character the complex pose (action, facing, what the hands are doing, visible emotion)
• Give the second character only a simple state: standing, crouching, back turned, seated
• Need both complex? Split it into two images, or put one in the foreground and one far away

Rule 2: Write “where the camera stands,” not just a focal length

The trap: I wrote “85mm telephoto, compress the distance between the two people.” It had no effect, both times. What worked was “the camera stands outside the door, looking in through the doorway,” first try.

Why: models understand spatial relationships far more reliably than optical parameters. “85mm” is just a number to the model. “Standing outside the door looking through the frame” is a scene it can see.

What to do: write the camera as three things.

1. A physical position: by the door, the far end of the table, the other side of the car, flat on the ground, a stair landing
2. What blocks the view: anything cutting into the foreground (door frame, railing, car hood, branches)
3. Height: ground level, waist, chest, eye level, looking down

A focal length can go last as a supplement, but never write only the number. If you really want the compression of a long lens, describe the effect itself: “the two figures look pressed together, the wall behind them flattened and pushed close.”

Rule 3: Write the visible result, not the action in progress

The trap: I wrote “the child is half-turning to look back,” and the child did not look back. Rewritten as “the child’s profile is visible, looking back over the left shoulder,” it worked first try.

Why: the model draws one instant, not a process. “Turning” is the middle of a motion and the model does not know which frame to draw. “Profile visible, over the left shoulder” describes what that instant looks like.

What to do: when you write an action, ask “if I pressed the shutter, what would this frame show?” Then write that.

Rule 4: Negative instructions do not work

The trap: I wrote “no sky visible,” and the moon appeared anyway. I confirmed the same result on a completely different model, so it holds across models.

Why: most image models handle “not X” poorly, and some become more likely to draw X because you mentioned it. There is also a technical reason: on some “fast mode” models (turbo, lightning, fast), the negative prompt does nothing at all, because their guidance scale (CFG) is set to 1, and at that setting the negative prompt does not take part in the computation.

What to do:

• To control composition, describe it positively and fill that spot with something else. If you do not want sky, write “the top of the frame is filled with the eaves and hanging straw”
• Do not expect “no XXX” to remove an element
• For style words you do not want, the best move is to leave them out of the positive prompt, not stuff them into the negative one

Rule 5: Every light must be explainable inside the frame

What to do: for every bright spot, you should be able to point to where the light comes from: a window, a lamp, fire, the moon, headlights, a warning light, a screen, a projector.

Avoid:

• Rim light with no source (a character’s edge glowing for no reason)
• Blue-and-orange with no reason (unless the frame really has a blue source and an orange source)
• Crushed blacks and blown whites (keep detail in the shadows)

A bonus move: drop the exposure by one stop and keep shadow detail. An evenly over-bright image is the classic AI look.

Rule 6: Leave one “unfinished message” (the best return on effort)

Across the six images, this kind of element had the highest hit rate, and the cost is one extra sentence in the prompt.

What it is: a detail that hints the story is not over. What worked in testing:

• One shoe left unpicked on the grass at the threshold
• A window still lit in a distant house
• A phone dropped in a puddle with its screen still on
• A figure standing in that lit window (the strongest one: it turns “they are leaving” into “someone is watching them leave”)
• The profile of someone looking back

Why it matters: it is the line between “a pretty picture” and “a frame from a film.” A picture with no unfinished message is a poster, not a story.

Before you write, answer four questions

Answer these four before writing any prompt. If you cannot, the scene is not thought through yet, and what you write will be empty.

1. Story: what happens in this instant?

• What happened a second ago? What is happening now? What might happen next?
• Why is this person here? What decision are they about to make?
• What is the unfinished message?

Discipline: one main thread, one secondary thread. Everything else can be plain, dark, soft, blocked, or empty. Give only 2–3 concrete details per image. A pile of nouns is the first source of the AI look.

2. Light: why is this moment lit?

• What is the source? (It must make sense inside the frame.)
• Time, weather, direction, hardness
• Is detail kept in the shadows?

3. Space: how does the frame run from foreground to background?

• Is there something in all three layers: foreground, midground, background?
• Does the foreground occlusion have a story reason rather than pure decoration?
• What frames the shot: a doorway, window opening, row of columns, corridor, car body, branches
• Scale of the figure: a small person in a big space, or an intimate close distance?

4. Camera: where is the viewer standing?

• Which physical position is the camera in? (See Rule 2.)
• What blocks it?
• Height?

One more question people forget is color: a main color plus a secondary color plus one accent, kept to 2–3 relationships. Low saturation, deeper shadows that keep detail, with color coming from the scene’s actual light sources.

Prompt assembly order

After the four questions, put one paragraph together in this order (120–180 words recommended, under 200):

1. Story instant (who, doing what, the unfinished message)
2. Foreground: what stands in front
3. Midground: the subject
4. Background: the far layer
5. Camera position (where it stands, what blocks it, height)
6. Light (where it comes from) + exposure (down one stop, keep shadow detail)
7. Color (main, secondary, one accent)
8. Texture / grime (pick only one)
9. Aspect ratio

Why this order: story before style, staging before color. Models tend to weight the front of the prompt more, so putting the story first keeps it from being drowned out by texture words. Pick just one kind of grime (dust and dry scratches, rain and wet reflections, smoke, faded cloth and peeling paint). Mixing them all is the classic AI look.

Traps you will hit

Trap 1: Thinking more adjectives help

Words like highly detailed, intricate, epic, masterpiece, 8k, ultra realistic do not improve the image. They make the model spread an even, premium finish across the whole frame, which is exactly where the AI look comes from. Fix: leave them out and replace abstract quality adjectives with a concrete scene description.

Trap 2: A long negative prompt that does nothing

With a fast model such as turbo, lightning, or fast, the negative prompt may have no effect at all (it does not take part when CFG is 1). Fix: check the model setting first. If CFG is 1, skip the negative prompt and instead keep unwanted words out of the positive one.

Trap 3: Switching models, then finding out it is not commercial

This is the most expensive trap. Open weights do not mean commercial use is allowed: some flagship versions are non-commercial and only the small models are permissive; some have revenue or traffic thresholds; some require outputs to be labeled as AI-generated; and many high-star GitHub projects have no license at all, which means all rights are reserved by default and you cannot use them commercially. Fix: for any model going into a for-profit use, read the actual LICENSE text before you start, not a second-hand summary.

Trap 4: Solving it by rerolling

Rolling the same prompt 20 times is gambling, and each roll costs money or time. Fix: lock the seed, change one variable at a time, and see which change really works. Three directed experiments beat twenty blind rolls.

Universal rescue: how to ask an AI when you are stuck

When you are stuck at any step, copy the block below to an AI assistant and replace what is in [] with your situation:

I am generating images with [tool name, e.g. ComfyUI / Midjourney / Stable Diffusion WebUI],
my operating system is [Windows / macOS], I am a beginner, please tell me step by step what to do.

The effect I want is: [describe the image you want]
My current prompt is: [paste the full prompt]
What actually came out: [describe what is wrong]

Please help me:
1. Decide whether it is a prompt problem or a settings problem
2. If it is the prompt, give me a corrected version directly and explain what you changed and why
3. If it is a setting, tell me where to click and what value to use

Do not just give me the theory. Give me steps I can follow directly.

If it is an error message, paste the full original text (do not just write “it threw an error”) and add: “This is the full error message. Tell me the most likely cause and the first thing I should check.”

One-line summary

Do not fix a disobedient image with more adjectives. Lock the seed, change one variable at a time, and find which two descriptions are fighting. If you can remember only one of the six rules, remember this one: leave an unfinished message. It costs one sentence and gives the biggest payoff.