How Do I Write Richer AI Image Prompts?

Updated 2026-10-27·3 min read·Rendering

Why adjectives don't work

We learned this twice. When we wanted a model to list more props from a screenplay, deleting the word "significant" from "list every significant prop" changed the count by zero. Telling it to walk through the screenplay scene by scene took the count from 12 to 19.

The same thing happened with prompts. Asking for "richer, more cinematic" prompts didn't move anything. What moved them was spelling out the decisions behind a cinematic frame.

An adjective names a quality the model believes it's already delivering. A specific decision it hasn't made yet is something it can see is missing.

The classic anti-pattern is a canned tail like "cinematic lighting, sharp focus, film grain" pasted onto every shot. It's identical on every frame, so it says nothing about any of them.

The checklist

For each shot, decide:

1. Frame: shot size, camera height and angle, all three. "Medium close-up, low angle, camera at chest height."
2. Move: one camera move, or "locked off." (For a still, this tells you what the next frame wants.)
3. Light: its source, its direction and its quality. "Single desk lamp, frame left, hard."
4. What the light does: "throws his shadow up the wall behind him."
5. Surface: what things are made of and how they look. "Chipped enamel mug, scratched oak desk."
6. Air: what's between the lens and the subject. "Cigarette smoke hanging in the lamp beam."
7. Sound, for video: what we hear. "Rain on the window, a clock ticking."

Before and after

Before: "A detective at his desk at night, cinematic lighting, ultra detailed, moody."

After: "Medium close-up, low angle from desk height. Locked off. A single green-shaded desk lamp frame left, hard light, throws his shadow large on the wall behind him. Scratched oak desk, a chipped enamel mug, case files fanned out. Smoke hangs in the lamp beam."

The second one makes each of those decisions for the model, and the first leaves them to chance.

Two traps

• Don't paste your project's style into every prompt. If your look, grade or medium is set once for the project, restating it per shot creates copies that go stale the moment you change the setting.
• Pick a model that will draw the beat. In our tests, some image models refused intense moments outright. Nano Banana and GPT Image 2 both blocked a two-person slap, while Seedream drew it, and gave the strongest emotion on that violent beat. A rich prompt can't help a model that won't render the shot.

In ShowMaker

Cast, locations, props and wardrobe are @references, so the same face, room and prop are sent with every shot. Your prompt doesn't need to re-describe them and can spend its words on frame and light instead. Break scenes into shots with size and camera move on the storyboard, or start from the AI storyboard generator. If a frame gets refused, see why AI video prompts get blocked.