Your prompts are describing the mood. They should be directing the shot.

There is a pattern in the prompts that do not work, and once you see it you cannot unsee it. They are full of words like cinematic, moody, atmospheric, dramatic, beautiful, haunting. These words describe how the result should feel to someone watching it. They say nothing at all about what should appear in the frame, how the camera should be positioned, what the subject should be doing, or how the light is physically falling. The model has no feelings to match against. It has training data. Adjectives that float above the image, untethered to anything concrete, give the model permission to guess, and it will guess the median of everything it has ever seen tagged with that word.

The fix is not to write longer prompts. It is to write prompts that work the way a director's instruction works: position, subject, action, light source, lens behaviour, duration. A director on set does not tell the camera operator that the shot should feel melancholy. They say where to put the camera, what the lens should be, what the actor does, and how long it runs. Your prompt is the only communication channel you have with the model. Every word in it either narrows the possibility space toward what you want, or it does not. This tip is about replacing the words that do not narrow anything with words that do.

The three categories of useless adjective

Before rewriting anything, it helps to sort the offenders into groups.

Mood adjectives tell the model how you feel about the result: cinematic, epic, beautiful, haunting, dreamlike. They are marketing copy, not direction.

Quality adjectives signal ambition without providing information: ultra-detailed, photorealistic, stunning, masterful. The model already has a quality dial; these words do not turn it.

Vague genre markers pretend to be specific: noir, vintage, indie film look. These can mean ten different things visually. If you want low-key side lighting from a single practical source, write that.

None of these categories needs to be banned entirely. A genre label can anchor a cluster of visual conventions, but only if you then follow it with specifics that define which version of that genre you mean.

The director's rewrite: a before and after

Here is a typical mood-led prompt:

```
Cinematic shot of a woman walking through a rainy city at night, moody and atmospheric, dramatic lighting, ultra-detailed, beautiful and haunting.
```

Now here is the same scene rewritten as direction:

```
Low-angle tracking shot, camera at knee height, following a woman from behind as she walks along a wet pavement at night. Shallow depth of field, foreground blur from rain on lens. Single overhead sodium streetlamp as key light, casting a hard downward shadow. Neon shop signage in background, out of focus, red and green. She walks at a steady pace, no hurry. No camera shake. 10 seconds.
```

Count the differences. The second prompt specifies camera height, camera angle, camera movement, subject position relative to camera, depth of field, a specific light source with a named colour temperature, a specific background element with named colours, subject behaviour, camera stability, and duration. The model now has a series of concrete decisions to honour. It cannot default to the median cinematic night scene because you have closed most of those doors.

The five slots every prompt should fill

Think of your prompt as having five slots. If any slot is empty, the model fills it with its own default, and defaults are average.

1. Camera position and angle. Not just wide shot or close-up, but where the camera is in physical space relative to the subject. Knee height. Eye level. Looking down from a second-floor window. Behind the subject's left shoulder.

2. Subject and action. Who or what, doing what, at what pace, with what body language. A man sitting is a default. A man sitting with his forearms on the table, looking down at his hands, not moving is a direction.

3. Light source and quality. Name the source: window light from frame left, a desk lamp pointing at the ceiling, fluorescent strip overhead, a candle at table height. Name what it does: casts a hard shadow to the right, wraps the face softly, underlit.

4. Lens behaviour. Depth of field, focal length character (compressed background, wide slight distortion), rack focus if the model supports motion prompts.

5. Duration and movement. How long. Does the camera move or hold. If it moves, how: slow push in, static, slow pull back.

What this changes in practice

The shift you will notice first is consistency across takes. When your prompt is specific, the takes that come back resemble each other more closely, which means you can select across them and cut without the viewer feeling a jolt between two clips of what should be the same scene. Mood adjectives produce high variance. Concrete direction produces takes that cluster around the same visual result, giving you actual coverage to edit with.

The second shift is that your prompt becomes a reusable asset. A prompt written in the director's format can be modified by changing one slot at a time, maintaining the rest, to produce a different angle or a different beat of the same scene. Mood-led prompts do not fork cleanly; they have to be rewritten from scratch each time.

Start with one scene you have already generated and rewrite its prompt using the five slots. Run three takes with the new prompt and three with the old. Look at the spread of results in each group, not the best result, the spread. That is where the difference lives.