Text to video AI: how to write prompts for Wan 3.0
By Hepifun Team
A prompt tells the model what to show, what happens, how the camera moves and how the scene looks. Most weak clips come from prompts that leave one of these out, or that ask for too much at once. This guide gives you a four-part structure, two complete examples and the common reasons a prompt fails.
The four-part structure
- Subject. Who or what is in the frame. "A street food cart," "a golden retriever."
- Action. What happens. "Steam rises from the pan," "runs across the beach."
- Camera. How the camera moves, or that it stays still. "Slow tracking shot," "static wide shot."
- Light and mood. Where the light comes from and how the scene feels. "Warm light from string bulbs," "soft orange backlight."
You do not need every part in every prompt. Use all four when you want full control.
Worked example 1
A street food cart at a night market, steam rising from the pan, the handheld camera slowly follows the cook, warm light from string bulbs. Sizzling sound and distant crowd chatter.
- Subject: street food cart, cook
- Action: steam rises
- Camera: handheld, slow follow
- Light: warm string-bulb light
- Sound: sizzling and crowd chatter
“A street food cart at a night market, steam rising from the pan, the handheld camera slowly follows the cook, warm light from string bulbs....”
Worked example 2
A ceramic teapot on a wooden table, steam rising, soft morning light from a window, slow push-in.
A short prompt like this works because it has one subject, one action and one camera move.
“A ceramic teapot on a wooden table, steam rising, soft morning light from a window, slow push-in.”
Why a prompt fails
These are common problems:
- Too many actions. "A dog runs, jumps, swims, and a car drives past." Pick one main action.
- Contradictions. "Static camera with a fast tracking shot" gives mixed motion. Choose one camera move.
- Vague subject. "A nice scene" gives the model nothing to hold on to.
- Real people. Prompts that ask for real public figures, or for a face swap, are rejected under our Content Policy.
- Text in the video. Written words often come out garbled. Keep text out of the prompt unless you need it.
Language
Prompts work best in English. We have not tested Indonesian prompts yet.
Length
Two to four sentences is usually enough. Long prompts with many details can make the motion less clear.
Test before you pay for 1080p
Write your prompt, generate at 480p (10 tokens for 5 seconds, Rp6,000), check the result, then move up to 720p or 1080p. Read how much an AI video costs per second for the full rates.
Ready to try? Write your prompt.