Free-form prompts are fine for one-off experiments, but they become hard to maintain when you generate many clips. After a few dozen renders I started writing prompts as if they were a small record with named fields.
subject: a paper lantern hanging over a canal at dusk
action: lantern sways, reflections ripple below
camera: slow push-in, eye level, shallow depth of field
style: warm morning light, soft shadows
duration: 10s
avoid: text on screen, extra objects
Why this helps:
- Diffs are readable. When a render goes wrong you change one field, not a paragraph, and you can see exactly what changed between versions.
- Fields map to inputs. The subject field can often be replaced by a reference image; the camera field can be replaced by a reference clip. The template shows which parts are better shown than told.
- Reuse across a series. Keep style and avoid constant, vary only subject and action, and a set of clips starts to look like it belongs together.
I flatten the record into one sentence per field before sending it. In testing with Wan 3.0 (AI video generator), which runs the Wan 3.0 model, this format worked well alongside its other inputs: up to three images, a reference video for motion, first/last frames, audio and documents. Output length went up to 30 seconds per run. Each attempt shows its credit cost in advance, so field-by-field experiments stayed cheap, and there is a free tier without a card.
The template is not magic, but it turns prompting from guesswork into something closer to configuration.