From Rough Idea to Cinematic Video: Character Sheets and JSON Prompting

Spending hours on long, complex prompts and still missing the frame you pictured may not be a skill problem at all. The host shows short films he made with zero prompt-writing skill and reduces his claim to one sentence: describe the raw idea in plain words, and let a separate tool write the professional prompt. The whole video is a workflow that tests that claim across two scenes.

But no matter how good the prompt gets, one unlocked variable ruins everything: the character has to look like the same person in every shot. The answer is the character sheet , a reference page locking the same face and outfit from several angles. InVideo's consistency guide states the same rule: prepare a multi-angle sheet up front, lock the outfit into it, and attach the same page as a reference to every generation. That turns visual consistency from luck into a system.

Sheets locked: two identities for indoors and outdoors

The production runs inside Higgsfield, an all-in-one platform the host openly names as his sponsor. According to Higgsfield's own guide, GPT Image 2.5 ships in two variants: fast Flare for everyday work and premium Sunburst for precise edits. The host deliberately picks Sunburst for the character sheets. The platform's video page supports that choice: models such as Veo, Sora, Kling and Wan live in one workspace, with continuity controls like first-frame and last-frame locking served under the same roof.

The first scene plays in a mountain cabin: the character pours his coffee inside, then heads out to chop wood. The host builds the indoor sheet first, attaching a real photo of himself as the identity reference and asking for four panels : front, three-quarter, side and a face close-up, at 16:9 and 2K resolution. The result surprises him because the generated face genuinely looks like him, right down to small proof-of-realism details such as slightly chapped lips.

The outdoor sheet uses a cleverer trick: two images doing two separate jobs. The first is again the host's real face, locking identity. The second exists only for texture matching , a frame already known to look like a real photograph, used so the model copies skin and fabric texture from it. The cold-weather sheet therefore keeps the exact face of the indoor sheet while holding photographic realism. Together the two sheets form the reference set that will carry the whole scene.

Plain sentence or structured prompt?

The video tests run on Seedance 2.5, ByteDance's video model. The lazy route goes first: the scene is written out as it sits in the head, set to 16:9, 30 seconds and 1080p, both sheets attached, then generated. The result looks decent but carries two flaws: odd cuts that break the flow and a log error, with the character striking the big trunk that should serve as the chopping block. According to ByteDance's Seed page, the model builds 30-second narratives in a single run and can extend twice; TechNode's launch report adds that one input can carry 30 images, 10 videos and 10 audio clips as references. So the model is strong, but a plain prompt leaves nearly every real decision, camera, light and pacing, to it.

The second route is the video's real star: a free tool called videoprompt.studio. The rough idea goes into a box in plain words, a shot type is picked, the target model stays on Seedance, and the output mode is set to JSON. One critical rule is written straight into the plain idea: the scene starts inside and moves outside, so the outfit must switch at exactly the right moment. The tool returns a paste-ready structured prompt split into fields such as lighting and camera movement. In the words of Neurosignal's 2026 guide, such tools translate between imagination and model: they turn casual description into the format video models can actually process.

The structured prompt goes in with the same sheets and the difference shows at once: cabin lighting and steam on the window look far more realistic. But a read-back catches two gaps: the dog and some motion-and-audio details never made it into the JSON. The lesson is harsh: whatever the prompt omits, the model never sees. After adding the missing bits and rebuilding, the third version comes alive, with the character picking the log up off the ground. SeedanceTool's model comparison explains why: Seedance responds best to prompts written with explicit camera moves, lighting setups and duration data. Template and model matched.

The hard test: a storm where everything changes at once

The cabin scene was the easy version because almost nothing changes state on screen. The second scene is the opposite: a night storm where the door swings from shut to open, the lantern from dark to lit, and the storm itself from full force to calm. The character sheet carries over from the first scene with only the wardrobe swapped to a night look, while the dog is built from scratch, from a written description into a four-panel sheet. Both the plain prompt and the raw structured prompt stumble: the character idles by the door, and he lights the lantern through the glass. The diagnosis is sharp: these prompts say what happens but never when , so the order of changes drifts.

The fix is to stage the prompt scene by scene: the power cuts first, then the door blows open, the character forces it shut, the lantern only catches near the end, and the storm clears at the finish. That ordered version lights the lantern properly and the movement turns natural. The closing lesson is about shuttling between platforms: separate tools mean separate credits, and the host argues that keeping the chain under one roof saves time and money. Read that with a sponsor discount, but the method itself is tool-independent: lock the sheet, structure the prompt, schedule the change.