Writing for AI video models is its own skill. The same prompt that produces a forgettable clip on one try can produce something striking on the next — the difference is almost always in the wording. Here are five patterns that consistently work.
1. Name the camera, not just the subject
"A cat" is a noun. "A cat, low-angle tracking shot, golden hour, 35mm" is a scene. Modern text-to-video models reward language a director would use: shot types (close-up, wide, dolly, drone), lighting (soft, harsh, golden hour, neon), and lens hints (35mm, anamorphic, shallow depth of field).
Try replacing every noun in your prompt with the same noun plus three production words.
2. Lock the style in the first ten words
The first words of your prompt carry the most weight. If you want a specific look — cinematic, anime, claymation, documentary — say it first. "Cinematic, slow motion, a runner crossing the finish line" outperforms "A runner crossing the finish line, cinematic, slow motion" by a noticeable margin in our tests.
3. Show motion, don't just describe a frame
The model decides the motion based on your verbs. "A boat on the water" gives you a still. "A boat slicing through choppy waves" gives you a moving shot. Use verbs that imply movement direction — drifting, racing, falling, rising, spinning — and the camera tends to follow.
4. Bound the imagination
Long prompts aren't always better. Models trail off when you stack too many adjectives. A clean prompt with one strong subject + one clear action + one stylistic note (three short phrases) tends to outperform a four-sentence essay. Try writing it the long way, then cutting until only the load-bearing words remain.
5. Iterate on the same seed idea
If the first generation almost worked, don't change everything — change one thing. Same prompt, swap "evening" for "dawn". Same prompt, add "from below". Treat each generation as a draft. Three iterations of the same idea is usually more productive than three completely different prompts.
These five rules don't replace experimentation — they accelerate it. The fastest way to find what works is to generate, watch, and adjust one word at a time. Try the examples for prompts that already pass the test, or start from scratch with the patterns above in mind.