Directing Multi-Shot AI Narratives Using Structured Reference Sheets and Draft Mode
AI filmmakers can avoid costly generation errors by pairing structured prompt formulas—such as 'start scope' and 'lock' formatting—with draft-mode testing and persistent asset sheets. Creators demonstrate that using language models like Claude to write multi-shot prompts alongside rigorous asset preparation cuts credit waste and preserves character consistency across extended sequences.
Rourke
Top learnings
Always separate multi-scene scripts into distinct generation blocks and test initial motion in draft mode to conserve generation credits.
Prepend prompt references with unified style formulas and explicit start-scope tags to prevent character drift and background warping.
Incorporate a two-second quiet audio buffer at the end of generated clips to prevent dialogue lines from clipping during editorial assembly.
How Structured Prompting and Claude Orchestrate Multi-Shot Sequences
Writing individual video prompts manually often results in generic motion and unpredictable visual quality. Creator Thomas Lundström and collaborator Caleb demonstrate that using an AI assistant like Claude to ingest character sheets, location references, and script requirements yields highly detailed, cinematic shot lists. By instructing Claude to structure prompts with a style prefix followed by scene descriptions and timecoded cuts, directors maintain strict creative control.
Building on this multi-shot approach, Kiko implements a rigid prompt formula split into 'start scope' (defining the exact purpose of each attached reference image), 'lock' (specifying camera position and character actions), and a clean ending outlining negative constraints. This systematic handoff ensures that downstream video engines map identities correctly without confusing props or background geometry.
ReelStack suggests adopting this structured prompting workflow for episodic pre-production to standardize multi-scene handoffs. However, filmmakers must trade away spontaneous generative variety, as hyper-specific prompt locking can occasionally produce rigid or overly formulaic motion vectors if the underlying reference assets lack sufficient variation.
Read source excerpts 2
references, all into Claude. You then describe the scene you want with Claude, and at the end of it, ask it to make you a SeaArt 2.5 prompt with these changes added to the
Rourke
personally like to use my short-list director skill for this, which structures the prompt with a style prefix, then it describes the scene and the characters, and then lastly,
Thomas Lundström
Why Asset Standardization and Draft-Mode Testing Prevent Credit Burn
Generating full-resolution AI video clips without preliminary testing is a primary driver of wasted production budgets. Thomas Lundström notes that rendering initial video tests in a lower-cost draft mode allows directors to audit camera trajectories, blocking errors, and prompt alignment before committing expensive hero credits to final renders.
To ensure these drafts succeed, character and location assets must be meticulously prepared. Creators emphasize establishing standardized reference sheets—such as pairing facial close-ups with front and back full-body views—while deliberately stripping out stray text markings or complex visual noise patterns that diffusion models might misinterpret.
ReelStack suggests implementing a mandatory draft-and-upscale protocol across all commercial AI pipelines to reduce generation expenditure by over fifty percent. Note that draft modes frequently lower output sharpness and temporal stability, requiring careful visual verification before upscaling to final resolution.
Thomas Lundström Read source excerpts 2
to look AI. For the characters, I always use the same template, a close-up of the face and then a full body front and back, and the front one doesn't have a face. And to
Thomas Lundström
When we're satisfied with the cuts and transitions, we generate our first video using the assets and the prompt. And my first generation, I have started to do always in draft
Thomas Lundström
Practical Filmmaking Application: Assembling a Multi-Part Narrative Short
Applying these structured asset and prompting techniques to a complete project involves partitioning multi-scene scripts into manageable time blocks and maintaining continuous spatial overlaps. Kiko illustrates that attempting to generate an entire action sequence in a single pass introduces severe tracking failures and spatial drift. Instead, directors should split narratives into distinct time segments—such as a 15-second kitchen setup followed by a 21-second living room sequence—and save the exact final frame of each preceding clip to use as the starting keyframe for the next generation.
Furthermore, when dealing with dialogue and audio cues within these generated clips, Kiko emphasizes that dialogue often clips abruptly at generation boundaries. Directors must program a two-second quiet audio buffer at the end of each shot and integrate dedicated voice synthesis tools like ElevenLabs in post-production to preserve vocal clarity.
ReelStack suggests utilizing this frame-matching and audio-buffering workflow to stitch seamless 30-second clips into longer episodic formats. The primary trade-off is the added editorial overhead required in non-linear editing software to trim boundaries and align secondary audio tracks manually.
Read source excerpts 2
kitchen and then a 21-second long generation in the living room. This way, we can be more specific with the cuts and vibe of each generation and C dance doesn't get confused.
Thomas Lundström
first frame of our next generation. So how do you connect the second 30- secondond clip without the scene jumping? That's key frame 4. We didn't write a prompt for it at all.
FRMWRKD-EXPLAINED
Operational Limits and Mechanical Failure Points in Video Diffusion
Despite rigorous pre-production asset prep and structured prompt formulas, generative video models remain susceptible to physical logic breakdowns and rendering artifacts. Kiko highlights that complex physical interactions—such as multi-subject combat, floating props, or characters interacting with uneven architectural edges—frequently result in visual glitches like figures hovering mid-air or background geometry warping unnaturally.
Additionally, varying model architectures introduce distinct financial and operational compromises. While premium engines like Seedance 2.5 deliver superior skin textures and precise physical adherence, they incur significantly higher credit costs per second compared to budget alternatives like Kling 4.0, which render faster but demand more re-rolls to fix tracking errors.
ReelStack advises directors to isolate high-risk physical stunts and complex prop interactions into practical inserts or low-resolution test passes rather than relying entirely on single-pass video generation. Filmmakers must weigh the rendering cost of flagship models against the time spent regenerating broken shots on budget platforms.
FRMWRKD-EXPLAINED Read source excerpts 2
Clling actually has pretty great quality, but when it comes to physics and errors, it makes way more mistakes. In our experience, Sea Dance saves you much more time because
FRMWRKD-EXPLAINED
from our rooftop scene. The villain on the right is literally floating in midair off the edge of the building. His scale doesn't match the roof. The red beacon pole grows
FRMWRKD-EXPLAINED
Key moments to explore
Optional deep divesWant to see a technique in action? Jump into the source videos. These AI-extracted timestamps may be approximate.
- 00:46 ↗Step one involves recording base footage with actors and stand-in props to give the generation model depth and interaction targets.Rourke · Seedance 2.5 Tutorial: Full AI Video-to-Video Workflow
- 01:45 ↗Use an LLM like Claude to co-develop concepts and convert plain language into structured prompts rather than writing complex prompt code manually.Thomas Lundström · 5 Mistakes Killing Your AI Realism in 2026 (Easy to Fix)
- 0:53 ↗Seedance 2.5 costs 210 credits compared to Kling 4.0's 67.5 credits on Higgsfield, but handles physics far more accurately with fewer re-rolls.FRMWRKD-EXPLAINED · Kling 4.0 Was Not What I Was Hoping For... (Honest Test)
- 01:19 ↗Step two requires building locations by taking a frame from the original video and prompting changes inside ArtCraft.Rourke · Seedance 2.5 Tutorial: Full AI Video-to-Video Workflow
- 03:21 ↗Standardize character reference sheets with a facial close-up alongside front and back full-body views to preserve identity consistency across AI generations.Thomas Lundström · 5 Mistakes Killing Your AI Realism in 2026 (Easy to Fix)
- 1:56 ↗Kling 4.0 supports generations up to 30 seconds per clip, doubling the 15-second duration limit of Kling 3.0.FRMWRKD-EXPLAINED · Kling 4.0 Was Not What I Was Hoping For... (Honest Test)
Put it into practice
Turn an AI film brief into a shot plan →Practical steps and checks before committing to production. Get the practical weekly briefing →The useful ideas in one email. Subscribe to keep learning. Need a filmmaker for your project? →Tell us what you want to make and submit a project brief.Go to the source
3 videos