Creative Space
Turn a Reference Video into Adapted Prompt Cards

Content

The agent obtains/analyzes the actual reference, chooses a model, and expresses transferable structure, composition, movement, light, and style in copy-ready prompt cards. Replace people, brands, and copy with the user's material. An optional representative-shot trial can be compared side by side in Flova.

Unlike the reference-video remake inspiration, this delivers prompts, not a finished film. The user can generate independently or continue to the full remake workflow.

Tools

  • Teardown — retrieve and read the full “Break Down a Reference Video into an Interactive Report” inspiration through Flova MCP by name, required: Obtain and analyze the reference.
  • FFmpeg, required for video: Frames, duration/aspect ratio, source comparison clips.
  • Flova: Current model parameters/reference requirements; optional trial generation.

References

  • , steps 2/4: per-shot/image fields, conversion into prompts, content not to copy.
  • , steps 3/4: Seedance 2.5, GPT Image 2, MiniMax H3 by reference type.
  • , steps 4–6: format, uploads/roles, reference wording, comparison.

Boundaries

  • Reuse structure, not source identities/content. Replace real people/celebrities, copyrighted characters, brands/logos, copy/dialogue/music with user-owned content or generic descriptions.
  • Mark illegible/inaudible details unconfirmed; do not guess.
  • Explain that prompts approximate the feel, not guarantee identical frames.
  • Explain trial quota use and obtain agreement before generation.

Workflow

Obtain → analyze/confirm → select model/replacements/confirm → write cards/confirm → optional trial → deliver.

1. Obtain

Use Teardown acquisition. If unavailable, request an upload/screenshot. For video inspect duration, aspect ratio, frame rate, audio using FFmpeg. Establish whole-video, selected-shot, or style scope; replacement product/person/theme; available user assets.

2. Analyze — confirmation

Use Teardown, then to complete each shot/image: shot size, camera movement, subject, action, space, lighting, style, sound, on-screen text, transitions. Also record shared elements/rhythm. Show the table for correction and confirmation of what to borrow.

3. Model and replacements — confirmation

  • Recommend a model using the cheatsheet with reasons; honor an already specified model.
  • Query creation_generation_models, detail="full": durations, aspect ratios, reference counts/roles. If unavailable, mark unverified.
  • Group source shots into generation calls according to supported duration/multi-shot capability.
  • Agree replacements: user character photos/generic people, user brand/copy/dialogue; retain only music style, not source music.
  • Show model, groups, replacement list before writing.

4. Cards — confirmation

One card per generation, following and model-specific structure. For references specify responsibilities, exclusions, submission order/roles. Reuse identical descriptions/reference sets for cross-card people/products/scenes. State source shot and borrowed technique in each rationale. Show cards before trials or final delivery.

5. Optional trial

After agreement, choose the card best representing the source's feel and Quick-generate through the card reference. Extract the corresponding source clip/frame with FFmpeg for side-by-side viewing. Compare shot size, camera movement, light, palette, rhythm; explain similarities/differences. Revise and retry mismatches, changing one variable per round. Propagate approved wording to other cards.

6. Deliver

Deliver source analysis table; numbered prompt cards; model rationale and replacements. If tested, include comparisons, round changes, task IDs. Suggest the full reference-video remake workflow for a finished film, or the target model's prompt-optimization inspiration for further refinement.

Disassembled for writing prompts

Ordinary video breakdown is concerned with "Why is this video good-looking", while breakdown for writing prompts is concerned with "Which words should be written so that the model can generate a similar picture". This article explains which fields need to be disassembled for each shot or picture, how to change from disassembly to prompts, and what cannot be copied.

Video: Shot-by-shot teardown

First use the Teardown inspiration video breakdown to get the shot segmentation and time. Then take out each shot:

FieldWhat to dismantleExample
DurationStart and end time00:03.2–00:06.8
Scenery and perspectiveLong shot to close-up, looking straight up, looking down, looking upMedium shot, slightly lower camera position and looking up
camera movementDirection, speed, following object; whether to take one shotSlowly move to the right while panning to keep the character centered
SubjectVisible characteristics of the person or product: age, hairstyle, clothing, material, colorFemale in her twenties, short hair, beige windbreaker
ActionStarting state → Process → Ending stateSitting and looking out the window → Getting up → Walking to the door and looking back
Space and environmentPlace, time, weather, front, middle and back scenesOld-fashioned apartment living room in the evening, rain outside the window
Light and colorMain light direction, color temperature, contrast, color adjustmentCold light outside the window, a warm desk lamp indoors, low saturation
Style and mediumRealism, animation, advertising, cinematic feel; textureFilm grain, shallow depth of field
SoundDialogue (original text, speaker), ambient sounds, sound effects, and when music comes in and outThe sound of rain continues; the door axis creaks when the door is opened
Screen textOriginal text, location, styleNone
TransitionHard cut, dissolve, occlusion transition, matching editingocclusion transition before the character walks through the shot

Also note down the unifying elements of the entire video: recurring characters, products, scenes, overall tone and rhythm. These should be consistent on each cue word card.

Picture: Completely dismantled

FieldWhat to dismantle
PurposePosters, product images, key visuals, illustrations, character settings
SceneEnvironment, time, space level
SubjectAppearance, material, posture, action, relationship with other objects
CompositionScenery, perspective, focal length, subject position, white space, aspect ratio
Light and colorKey light, fill light, shadow, color temperature, main color
Style and texturePhotography, illustration, 3D, graphic design; key materials
Screen textOriginal text of each paragraph, position, level, font tendency, color

From dismantling to prompts

  • Merge first and then split: When there are many video shots, don’t take one shot at a time. Combine several consecutive shots of the same scene and the same subject into one shot, and decide how to combine them based on the duration supported by the target model and the multi-shot capability.
  • Write something visible: Replace the adjectives ("high-end" and "good atmosphere") in the disassembly with specific light, material, composition and movement.
  • Reference assets instead of description: When the user has his own character photos and product pictures, use reference assets to be responsible for the appearance, and only write what it is responsible for in the prompt.
  • Write according to the structure of the target model: See .

Cannot be copied

What is reproduced is the structure, composition, camera movement, rhythm, light and style, not the content of the reference video:

  • Characters: Real people, celebrities, and copyrighted characters in the reference video, do not include prompts to copy their looks. Use the user's own character reference, or write a general description (age, hairstyle, clothing).
  • Brand and Logo: The brand name, logo, and packaging text of the reference video are not copied and replaced with those of the user.
  • Screen text and dialogue: The copywriting and dialogue of the reference video are not copied, rewritten according to the user's content, or left blank for the user to fill in.
  • Music: Do not specify a reference video song, only describe the style, instrument, mood and rhythm.
  • Unique creative settings: The reference video has a particularly recognizable creativity (an exclusive character, a unique visual logo), reminding users to replace this part with their own creativity.

The content of the reference video can be faithfully recorded in the disassembly document, and only the replaced content will be placed in the prompt card.

Key points in model selection and writing

After dismantling the reference video, you need to decide which model to use to reproduce it, and then write prompts according to the structure of that model. This article gives selection suggestions and writing points for three commonly used models. When in-depth diagnosis and repeated optimization are required, use the prompts corresponding to the model to optimize inspiration.

The duration, aspect ratio, resolution and number of reference assets actually supported by the model are subject to the query results of creation_generation_models. When the user specifies another model, follow the Flova Skill's "Write Generation Request" and explain that the writing method in this article may not be applicable.

How to choose

What is reference videoRecommendedReasons
Pictures: posters, key visuals, product images, infographics, especially accurate textGPT Image 2Strong text rendering, layout and precise editing
Video: long narrative, continuous action, multiple reference assets (characters, products, scenes, actions), realistic advertisingSeedance 2.5Multi-modal reference and long-term line strength, support editing, extension, transition
Video: A multi-shot short film where images, dialogues, ambient sounds, and sound effects are produced togetherMiniMax H3The sequence of shot language and sound is clearly written

When in doubt, tell the user the choice between the two candidates and let the user choose.

Seedance 2.5

Structure: Reference assets responsibilities + one sentence summary + specific timeline + global preservation constraints.

  • One-sentence summary = subject + location + core event + subject style + special shot language.
  • The timeline follows "Opening → Development → Climax → Ending" to write images, scenes, camera movement, action, dialogue and sound.
  • For each reference asset, clearly state what will be transferred and what will be excluded. For example, "The reference video only provides action rhythm, and does not inherit characters and backgrounds."
  • The number of shots, action density and dialogue length should be matched with the duration; only one core action should be left in the short video.
  • Only write the continuous path and camera movement from shot to shot, do not mix in cut shots.
  • Known limitations: The aspect ratio of the reference image must be between 0.39 and 2.50. If it exceeds the aspect ratio, it will be rejected; images that are too wide and too high will be padded first.

GPT Image 2

Order: Purpose → Scene → Subject → Composition → Light and Color → Style and Texture → Screen Text → Reference Image Responsibilities → Keep Items → Don’t.

  • Purpose is placed at the beginning. Write content that is visible and verifiable, and don’t use words like “high quality”.
  • Use quotation marks to give the original text of the screen text, indicate the position, level, font preference, color, and "only appear once, no other text will be added."
  • Assign a responsibility one by one with multiple reference pictures, and then explain how to combine them.
  • When editing, list the preservation constraints first, and then write the modified items; only one variable is changed in each round.

MiniMax H3

Structure: Reference assets responsibilities + one sentence core idea + global characters, scenes, style and sound settings + shot or event timeline + preservation constraints.

  • Write a complete sentence of "scenery + subject + action + camera movement + dialogue or sound" for each shot.
  • Retell the characters, costumes, props and scenes for each cut.
  • In the dialogue, clearly describe the speaker, language, emotion, inside or outside the picture; match the length with the shot.
  • When there is no dialogue, indicate that the character remains quiet, his mouth is naturally closed, and non-speaking actions are arranged.
  • Ambient sounds, sound effects, and music are written in the order of events when they enter, weaken, and stop.

Common to the three models

  • Reference assets are only responsible for one or two things, do not write "reference all assets".
  • Write identification anchors for people and products and repeat them across shots.
  • Negative constraints should be few and accurate, and positive and negative requirements should not conflict with each other.

Prompt Cards and Flova Handoff

These inspirations deliver prompt cards, not images/videos: one per image or shot. Cards can be used directly for generation; this reference covers writing, Quick generation, and validation.

Card format

Include all fields; write “none” where absent:

FieldContent
IDE.g. P01, mapped to the requested image/source shot.
Model/capabilityE.g. Seedance 2.5 multimodal video; GPT Image 2 editing.
PurposeOne-sentence intended use.
PromptComplete copy-ready text in one code block.
ReferencesOne row each: filename, submission order, role below, responsibilities, exclusions.
ParametersDuration, aspect ratio, resolution, etc., from creation_generation_models; otherwise “unverified.”
InvariantsUnchanged people, products, text, composition.
RationaleTwo/three sentences linking choices to model-writing rules.
Acceptance checksE.g. identity, verbatim text, absence of teleporting motion.

+Write in the target language. Chinese-user requests default to Chinese unless the model guide/user calls for English. Preserve dialogue/on-screen text verbatim; never translate/polish it.

Quick generation

+Follow Flova Skill's Quick/reference guidance:

  1. Query the model through creation_generation_models, detail="full"; verify parameters, reference slots/roles.
  2. Prepare references: supported aspect ratio/padding, crops, formats.
  3. creation_upload_ticket → transfer bytes to its URL → creation_upload_complete → file_ref. It expires in roughly one hour; upload again before submission when necessary.
  4. Assign each inputs role: first_frame starts matching that image; last_frame ends on it; reference guides appearance/style/content without exact matching; driving_audio drives lips/action; base_video is edited/continued.
  5. Submit with client_request_id; retain card/reference/file_ref/task-ID mappings.

Referencing inputs in prompts

  • Use ordinal natural language, e.g. first image controls face/clothing, second controls product. Match inputs order within each media type; images/videos/audio each count from first.

Other tools

+Remove Flova-specific markers/role instructions and use the destination tool's upload mechanism and parameter controls. Explain that these prompts target a particular model; changing models may change outcomes.

Optional comparison

+Explain quota use before proceeding. Generate original and revised prompts once each with identical parameters/references; show side by side. Change one variable per round (e.g. reference roles or timeline), record edits/task IDs/results, and deliver the record. Generation is stochastic: one comparison shows a tendency; repeat important findings once or twice before concluding.

Creative process

  1. FFmpeg

    FFmpeg

    Extract reference frames

    Inspect key frames for composition, action, and changes between shots.

  2. Flova

    Flova

    Generate test shots

    Generate visuals with the prepared prompts and compare them with the reference.