Turning a rejected client draft into a reusable AI video workflow
A client passed on an AI-generated video draft made in Higgsfield Cinema Studio with the Seedance model. The clips were too good to delete, so they became a public case study and a repeatable image-to-video production workflow.
- By
- Viron Gil Estrada
- Published
- Updated
- Read
- 6 min
In May 2026 I spent close to USD 200 in Higgsfield credits on a client video draft. The client passed on the direction. The clips were too good to delete, so I stitched the strongest ones into a reel, posted it on LinkedIn, and wrote down the workflow so the next project would not be a frankenstein. This is that workflow.
The stack that actually produced the clips
| Stage | Tool | What it did |
|---|---|---|
| Storyboard and prompts | Codex and Claude Code | Wrote the shot list, revised it, and generated the meta-prompts for every still and clip |
| Still frames | GPT Image 2.0 and Nano Banana Pro | Locked the look of each shot before any motion was generated |
| Motion | Higgsfield Cinema Studio 3.5 on the Seedance 2 model | Turned each still into a 5 to 10 second clip |
| Assembly | CapCut | Selection, pacing and the final cut |
The important structural choice is that stills come before motion. A still is cheap to regenerate and easy to judge. A clip is neither. Locking the frame first means every expensive generation starts from something already approved.
Four things I learned the paid way
Generation cost compounds. Every "one more try" is real money. Treat each prompt like a take on a film set, not a roll of the dice. Decide what you want before you press generate, and write it down so the coding agent can turn it into a precise prompt.
Client taste is not the same as work quality. A rejected draft can still be a portfolio piece and a learning artifact. The clips do not know they were rejected.
Stitching beats regenerating. Half of good AI video is editorial: knowing which three seconds of a five-second clip to keep. Cutting the strong seconds together produced a better reel than any single regeneration would have.
Live with some slop, and say so. There are warped hands and a frame where the physics gives up. Hiding that would make the case study useless. Six months earlier I could not have rendered any of it, and a year earlier I would not have known where to start.
The reusable workflow
- Brief in plain language. What is the product, who is watching, what is the one feeling.
- Have a coding agent write the shot list and a meta-prompt per shot, including negative constraints and the camera language.
- Generate stills for every shot. Approve or kill at the still stage. Never generate motion from an unapproved still.
- Generate motion at the model's native resolution and keep it there. Upscaling generated video degrades it; regenerate from scratch at higher quality instead.
- Cut for the strong seconds. Assemble, pace to the music, and stop.
- Log the spend per shot so the next brief can be priced.
The same skeleton now runs the AI content lanes I produce for Vibe Coders PH, with different models plugged into each stage as they improve.