Updated
Case evidence: Read the source breakdown report · View case frames and shot timings
Based on a published tool report. Production classifications are analytical judgments; suggested steps are not the original project or a verified recreation.
One shot, one structured paragraph
Here is the prompt the tool generated for shot 3 of the dialogue case: “Medium shot, eye level, the woman sits in the armchair left of center. She blinks once, then looks toward frame left and says calmly and firmly 「封面顺序反了」, with a small nod, both hands holding the mug throughout. Locked-off camera.” A shared first line sets format, light and references (vertical 9:16, low-light night interior, soft light, the character taken from a reference image). Framing, placement, action from start to end, the spoken line and the camera each get one clause. That structure transfers directly to Veo-style prompting.
Describe the audio, do not improvise it
Veo 3 can generate ambient sound and effects with the clip. Write the soundscape you observe — room tone, a door, footsteps — as part of the prompt. Keep spoken lines short inside the prompt, and decide early whether narration will be generated or recorded. In the source report, the background sound jumps at every cut, the analysis leans toward the voice having been generated with the picture (65%, unconfirmed), and no music bed was found. That matches the native-audio route Veo offers; if you expect to change the script often, recording the lines separately keeps edits cheap.
First-frame images keep characters consistent
When the same character appears across shots, fix the look once and reuse it. The dialogue report writes one first-frame image prompt per set — the living room for shot 2, the study for shot 3 — covering subject, wardrobe, seating, background props, light and framing, and reuses those frames for shots 5, 7 and 9. Every video prompt also points to a character reference image. Consistency comes from the reference images, not from longer prose.
Timing and cuts stay outside the prompt
The dialogue video cuts about every two seconds, but no single prompt should try to fake that rhythm. The report generates shots 2, 3 and 9 as separate clips, groups shots 5 and 7 — same sofa set, separated in the edit by an interface insert — into one multi-shot unit, and rebuilds the cut rhythm in the edit. If your engine supports multi-shot calls, group only shots from the same setup, and keep the total length inside the per-call limit.
What we have and have not tested
The prompts quoted above are real tool output from the published report. What is not tested here: we have not run these prompts through Veo 3 generations, and per-call limits change between versions. Use the structure, verify durations and audio behavior on one shot, then scale to the rest.
A reusable analysis template
Shot # / duration Subject (reference image id) / wardrobe Framing / camera behavior Action: start → end Light and setting Audio: ambience + effects (+ short dialogue if any) Keep out: captions, cuts, transitions
Practice with a real report
Open the dialogue case and read the report’s generation units: four units covering five character shots, each prompt with framing, action and camera, plus a first-frame image prompt per set. Copy that structure for your own reference video.
Open the case and download its report →Continue your workflow
- Video to prompt: translate observations into instructions
- YouTube video to prompt: analyze a reference shot by shot
- Why your video prompt does not match the reference
- How to reverse engineer video prompts from a reference
- How to reverse engineer a video, shot by shot
- Sora prompts: write them shot by shot from a reference
- Get started in Claude Code or Codex