How to get specific camera moves out of any prompt

Most AI filmmakers write motion into their prompts as an afterthought. The subject gets a careful description, the lighting gets a sentence, and then somewhere at the end: "camera slowly pushes in." The model obliges in the loosest possible way, and you end up with a shot that drifts vaguely forward, or bobs, or does something that technically qualifies as movement but carries none of the intention you had in your head. When you cut it next to another shot with its own lazy drift, nothing coheres. The edit starts to feel like a sequence of screensavers.

The problem is that natural-language motion description collapses the physical specificity that a real camera operator works from instinctively. A focus puller and an operator share a shorthand built from decades of standardised vocabulary: axis, arc, focal length, starting frame, ending frame, speed in feet per second if needed. That vocabulary exists because imprecision on set costs money. It turns out it costs you something in prompts too, and the fix is to borrow exactly that vocabulary, even though no model is physically operating a camera. Observable behaviour in current video generation models shows they respond more consistently to precise cinematographic language than to impressionistic alternatives.

Break the move into three components

Every camera move in a real production has three parts: the mount (what the camera is sitting on and how it moves), the frame at the start, and the frame at the end. When you include all three in a prompt, you give the model much less room to improvise in ways you did not intend.

Instead of: `camera slowly pushes in on the woman at the table`

Try:

```
Dolly push, starting on a wide shot framing the woman small
in a large kitchen, ending on a tight medium close-up of her
hands around the mug. Slow, constant speed. Camera stays low,
shoulder height. No tilt, no pan, purely axial movement toward
the subject.
```

The words "dolly push," "axial movement," and the explicit start and end framing give the model a corridor to work inside. You will still see variation across takes, but the variance is much tighter.

Name the lens behaviour, not just the focal length

Focal length changes how a move feels more than almost any other single variable. A push on a wide lens compresses quickly and distorts the background. A push on a long lens flattens perspective and isolates the subject. Neither is right or wrong, but they say completely different things emotionally, and they cut differently against other shots.

Use observable lens language:

```
Long lens, compressed perspective. Shallow depth of field.
Background bokeh holds soft throughout the move. Subject stays
sharp. The camera does not zoom; the lens length is fixed and
the camera moves physically toward the subject.
```

The instruction "the camera does not zoom" sounds odd but it genuinely helps. Current models conflate zoom and dolly frequently. Making the distinction explicit in the prompt reduces that conflation in the output.

Use static shots as anchors

One of the best habits from real cinematography that almost no AI filmmaker follows: build your sequence around locked-off shots, and let the moves be the exception. A locked frame is the easiest thing for a generation model to hold consistently. It is also the most useful edit tool you have, because a static shot gives the audience somewhere to rest and gives you a clean cut-point.

Prompt your static shots with:

```
Locked-off camera. Tripod. No movement, no drift, no breathing.
Static frame, medium shot. Subject moves within the frame.
```

That last instruction, "subject moves within the frame," redirects the model's need to add motion toward the character and away from the camera. Without it, models frequently introduce a slow creep or a handheld sway even when you explicitly ask for stillness.

Plan your axis before you prompt

The 180-degree rule exists because crossing the axis in the edit is disorienting for an audience. It applies exactly the same way in AI footage. Before you generate a scene, decide where your axis is and write it into every prompt in that scene.

```
Camera remains on the south side of the axis throughout.
Subject faces screen-right. Eye-line maintained across all shots.
```

This is a two-line addition to every prompt in a sequence, and it is one of the highest-leverage habits you can build. It will not guarantee perfect consistency across takes, but it gives you a filter for which takes to keep: any take where the subject's eye-line crosses to screen-left gets discarded immediately, no matter how good the motion looks.

The before and after

With vague motion prompts, you spend most of your review time discarding takes that went somewhere unexpected and hoping the next one lands closer. With specific cinematographic language, you spend that time making finer choices between takes that all fall within the corridor you defined. The ratio of usable takes goes up. More importantly, when you cut those takes together, they share a coherent spatial logic that the audience feels even if they could not name it.

The next sequence you plan, write the axis down on paper before you open a single generation window. Then write the mount, the start frame, and the end frame for every shot. Treat the prompt like a shot card, not a description.