Seedance 2.5 · Cinematic AI Film Workflow with Claude

Quick Summary

• Begin with references and a story. ShotDeck provides searchable examples of lighting, framing, color, lenses, cameras and emotional tone, while the script defines what the audience should feel.
• Use Claude as a translator and continuity assistant. Give it the reference frame, available metadata, script and non-negotiable visual rules, then ask it to produce Midjourney or Seedance-ready direction.
• Create separate reference sheets for every recurring character, location and important prop before generating video. Consistency is designed in preproduction, not repaired at the end.
• Write performance at a human level: eye direction, lip tension, eyebrow movement, posture, pauses, delivery, accent, hand action and the sound made by physical contact all belong in the shot description.
• Generate in roughly 30-second story blocks when the model and budget allow. Longer blocks can preserve spatial continuity across internal cuts, but dense dialogue may still need smaller sections.
• Treat every generation as source footage. Review it, dictate time-coded corrections, regenerate selectively and mine only the strongest seconds for the final timeline.
• Finish like a filmmaker. Tighten pacing, use scene detection to find internal cuts, match color across clips, add captions, build sound effects around movement and choose or generate music around the edit's emotional beats.

Credit, scope and what this tutorial actually proves

This guide is based on Rourke Heath's embedded tutorial, published on September 25, 2026, and the complete transcript supplied with this article. Rourke demonstrates the full process around a short dialogue scene in a park, including failed generations, prompt revisions and the Premiere Pro timeline. Credit for the workflow, example film and Video Prompt Builder method belongs to Rourke Heath and GenHQ.

The tutorial is a creator demonstration, not a controlled benchmark. Its practical value is the production logic: references, script, ingredients, long-form generations, review and finishing. Model access, duration limits, pricing and platform interfaces can change, so confirm those details inside the service you plan to use.

One important correction to the demonstrated workflow: do not rename protected products or alter references simply to evade copyright filters. Build original props, clear the rights to supplied assets and respect platform safeguards. Inspiration can guide lighting, composition and tone without copying a protected character, product or finished frame.

The complete pipeline at a glance

1. Define the message. Rourke starts with a creative position about AI agents and finds a familiar object that can carry the argument. The Tamagotchi-like virtual pet becomes a metaphor for why effort creates attachment.

2. Build the visual language. Search ShotDeck for frames that communicate the right emotion, then record the framing, color, lighting, camera, lens and location information that matters to the project.

3. Design original locations and characters. Use Midjourney or another image model to turn the reference language into project-specific frames. Approve character, wardrobe, prop and location sheets before video generation.

4. Direct the sequence in writing. Break the script into timed beats that describe camera placement, blocking, dialogue, acting nuance, ambience and physical sound.

5. Ask Claude to format the production prompt. Provide the script, reference images and global-look rules. Split the result into manageable duration blocks rather than forcing a full minute of dialogue into one generation.

6. Generate, review and revise. Use time-coded feedback, preserve successful moments and regenerate only the weak sections. Longer source clips are raw material, not a final edit.

7. Finish in an editor. Assemble the strongest takes, remove dead air, match color, add captions, sound effects, music and a final audio mix.

1. Use ShotDeck to define a visual language

ShotDeck describes itself as a fully searchable cinematic image database. Its frames are tagged with categories that include cameras, lenses, framing, lighting, color, composition, locations and even the emotion on an actor's face. That makes it more useful than a loose inspiration folder when a project needs repeatable visual decisions.

Search for the emotion and situation before searching for a fashionable look. Save frames that make you curious in a crowded feed, then ask why they work. Note the subject scale, camera height, foreground layers, contrast, color palette, depth of field and direction of light. If equipment metadata is available, record it as a reference rather than treating it as a guarantee that an image model will simulate optics perfectly.

Create a small visual bible with three groups: global look, recurring location and shot-specific composition. The global look remains constant across the film. Location references define geography and atmosphere. Shot references can change with each dramatic beat.

ShotDeck states that its images are for reference, research and education. Do not republish its stills as your own artwork or assume that a reference image grants commercial rights. Translate the cinematographic principles into original material.

2. Translate the reference into an original Midjourney location

Rourke uploads a selected frame and its available metadata to Claude, then asks for a Midjourney prompt that captures the cinematic style, lighting, composition, lens language and tonal qualities. The same frame can be used as both an image prompt and a style reference, depending on how closely the new composition should follow the source.

A useful rule is to separate what must transfer from what must change. Transfer the lighting logic, palette, lens character, depth and mood. Change the people, set dressing, geography, story action and distinguishing design so the result belongs to the new production.

Iterate with concrete comparisons. Instead of saying that a result is wrong, tell Claude that the camera is too distant, the bench is too small, the key light is too frontal or the background separation is too shallow. Rourke also increases Midjourney variety when he needs a wider range of possible locations, then selects and upscales only the strongest direction.

Store the approved location frame beside the global-look references. It becomes an input to the video prompt, a continuity target during review and a reference for later scenes in the same world.

3. Write the story and direct the acting before generating

The script is the production's control layer. Rourke writes as if he were directing a set: a wide aerial view holds for two seconds, park ambience and birds establish the space, the camera cuts to a phone reaction, a character becomes frustrated, throws the phone onto grass and delivers a line with a specific emotional quality.

Performance direction should describe observable behavior. Replace 'he is upset' with the action an actor would perform: the shoulders drop, the jaw tightens, the left eyebrow rises, eye contact breaks, the lip quivers or a smile is suppressed. Add the timing of the behavior and connect it to the line or camera cut.

Give characters stable voice and background cues, but avoid stereotypes. The tutorial keeps the second character's Australian voice cue consistent across prompts. A production bible can also record speaking rate, pitch, warmth, confidence, vocabulary and emotional baseline.

The metaphor makes the short memorable. Rourke's argument is that a virtual pet stops feeling meaningful if an agent performs all the care on the owner's behalf. The same writing pattern works elsewhere: define the abstract message, find a concrete object or ritual that embodies it, then let the conversation reveal the idea through action rather than exposition alone.

4. Build character, prop and location ingredients

Create one approved sheet for each recurring character. Include a clean full-body view, useful angles, a close portrait, wardrobe, footwear and any details the story references. Keep body proportions and identifying features visible. Rourke demonstrates generating a character sheet from a face reference, then editing clothing while preserving the person.

Use a separate sheet for a hero prop. Show front, back, side, scale and interaction details. If a recognizable commercial product creates a rights or moderation problem, the correct production solution is an original design with its own name and visual identity, not an attempt to sneak the protected item through a filter.

Add the approved location frame and global-look reference. Before generation, the input package should answer four questions: who is present, what they wear, where they are and what important object they handle.

Rourke uses Artcraft as the generation workspace in this demonstration. Treat that as a creator choice rather than an exclusive requirement. The transferable method is to choose a platform that exposes the Seedance version and reference controls you need, then verify its current terms, duration, reference limits and commercial-use rules.

5. Turn the script into a Seedance prompt with Claude

Upload the character, location, prop and look references to Claude, then paste the full script. Rourke asks Claude to make the prompt as descriptive as possible within Seedance's demonstrated 10,000-character input limit. The first attempt fails because the dialogue and action need about a minute, but the model is being asked to compress them into roughly 30 seconds.

The fix is structural: split the script into two duration blocks and give each block enough prompt detail. If speech accelerates or lines are skipped, reduce the amount of dialogue again. Duration is a storytelling budget. Every line, pause, reaction, camera move and transition consumes time.

Rourke installs a custom Video Prompt Builder through Claude's Customize and Skills interface. Anthropic describes custom skills as focused instruction packages for repeatable workflows, and recommends clear activation descriptions, examples and incremental testing. Review any downloaded skill before enabling it, especially if it contains scripts or tools.

A strong prompt builder should enforce a stable order: global look, cast and references, duration and aspect ratio, time-coded shot beats, performance and dialogue, sound, continuity constraints, and a short list of prohibited changes. The skill should format the director's decisions, not invent the story in place of the director.

6. Generate long blocks, then mine the best moments

Rourke prefers roughly 30-second generations because internal cuts can preserve the characters' positions, eyelines and physical performance more reliably than several unrelated six-second renders. That does not mean every 30-second result will be usable. One block may contain a beautiful reaction, an incorrect fall, a long pause and a rushed final line.

Mining means treating that output like filmed coverage. Mark the usable reaction, line delivery, camera move or transition. Keep it even if the rest of the generation fails. Build the final sequence from the best moments across several attempts.

Review with time-coded language. Rourke plays the clip silently and dictates observations with Wispr Flow: move the lip bite earlier, change the five-second shot from close-up to wide, begin the chuckle immediately at seven seconds and alter leg position so the actor does not look frozen. Claude then rewrites the prompt with those corrections.

Wispr Flow is only the capture layer in this loop. Its official documentation describes a dictation tool that turns natural speech into cleaned text in the active app. Any accurate voice dictation or manual note-taking method can serve the same purpose. Always read the transcript before sending it to the model.

7. Lock the global look and voice

Place the same global-look block at the beginning of every generation prompt. Include the chosen camera and lens language, aspect ratio, color response, lighting direction, contrast, texture, depth of field and any rules about exposure or camera movement. Claude can derive that language from a reference frame and its metadata, but the final block should be reviewed and frozen by the director.

Rourke demonstrates a voice-reference technique by recording a short sentence that covers many English phonetic sounds, placing that recording over a black frame and exporting it as an MP4 reference. He then identifies the video reference as the target voice for the corresponding character. Treat voice cloning as sensitive biometric use: record your own voice or obtain explicit, informed permission from the speaker, and check the provider's current policy before use.

A voice reference does not replace dialogue direction. Keep the character's accent, pace, tone and emotional intention in the prompt. Review every line for identity drift, strange emphasis and mismatched lip sync. Regenerate or replace any take that does not meet the performance standard.

8. Finish the film in Premiere Pro

Import every promising generation, not only the apparent winner. Rourke overlaps dialogue with the opening image, removes pauses that weaken social pacing and cuts between performances from different generations. The final scene is an edit, not a single successful prompt.

Premiere's Scene Edit Detection can identify internal camera changes in a rendered clip and create cuts, subclips or markers. Adobe recommends reviewing the detected points after analysis. This is a fast way to expose the individual shots hidden inside a longer AI generation so they can be trimmed and reorganized.

Match the grade across all selected clips. Rourke uses small Lumetri adjustments to saturation and exposure, then copies the treatment consistently. The goal is not to rescue fundamentally mismatched footage, but to remove minor differences after the same global look has already guided generation.

Add captions for silent viewing, then design sound around motion and emotion. Rourke layers ambience, impacts, risers and several music cues within the first section of the film. Artlist documents its catalog as licensed sound effects for adding realism, transitions and impact. Check the license that matches the channels and client work you plan to publish.

When stock music does not fit the cut, Rourke gives Claude the time-coded scene structure and asks for a music brief, then uses that description in Suno. The useful principle is synchronization: define tempo, key, instrumentation, emotional arc and the exact moments where the score must rise, drop or resolve.

Production checklist before delivery

Story: the short has one clear idea, a concrete metaphor and a beginning, escalation and resolution. Visuals: characters, wardrobe, props, geography, eyelines, lens language and grade remain coherent. Performance: dialogue pace feels human and reactions begin at the intended beat.

Edit: every shot advances story or emotion, dead air is intentional, internal cuts are clean and captions are proofread. Audio: dialogue is intelligible, ambience supports the location, sound effects match physical contact and music changes are motivated rather than decorative.

Rights and disclosure: confirm permission for faces, voices, reference assets, music and brand elements. Keep generation records, licenses and model settings with the project. For client work, describe what was AI-generated and what was edited or licensed by a human team.