I will not shut up about my free Claude Skills for realistic video prompting

Summary

Joey demonstrates an end-to-end workflow for making a cohesive AI-generated short film: a Halloween-season parody horror project inspired in part by American Psycho. His central idea is that AI should assist the filmmaker, not replace their creative judgment. Strong results come from preparation, clear visual direction, iteration, and human-led editing.

Workflow and techniques

• Develop the idea yourself first. Start with your own concept, notes, and collaboration with teammates. AI can help brainstorm, but the creative vision should come from the filmmaker.
• Build a story bible when needed. The Story Bible Builder skill interviews you about characters and the fictional world, identifies unresolved details, and creates a reusable reference file. Bringing that file into later sessions can help reduce continuity errors and hallucinated character details.
• Plan the story’s length and structure.
  • Establish a target runtime so the story does not become too sparse or overstuffed.
  • Develop a logline, then a concise synopsis, then a detailed chronological beat sheet.
  • In the beat sheet, specify what each character is doing, feeling, and reacting to in each moment—not just what happens in the plot.
• Tag and prepare assets manually. Identify locations, characters, props, and background characters. Create different character versions when their appearance changes, and make reference images for important props so their size and design stay consistent.
• Use Claude skills to prepare visual prompts. Joey describes using Cast Kit for character designs and reference sheets, SceneCraft for locations, and Shot Caller to turn story sections into prompts. Supply the story bible, beat sheet, and asset references, then generate one manageable section at a time.
• Keep prompts focused and give scenes room to breathe. Avoid cramming too much action or dialogue into a single generation. Include pauses, movement, reactions, nonverbal cues, and interactions with the environment. If a prompt is too long or produces rushed performances, ask Claude to simplify it.
• Build location references and maintain scene geometry. One suggested method is to animate a location image with a slow 360-degree camera turn, review several low-resolution versions, and use a useful frame as a reference. To continue an existing shot, use a screenshot of its final frame as the reference for an extended generation. Small contrast or color differences can be corrected in editing.
• Direct the camera with specific language. Techniques discussed include:
  • Handheld movement with subtle operator motion
  • Crash zooms
  • Dolly zooms
  • Whip pans
  • Dutch angles for unease or chaos

Joey also recommends adapting the prompting approach to the intended format—for example, a cinematic look with motion blur versus an iPhone-style look with sharper, higher-frame-rate footage.

• Test cheaply before rendering at high quality. Draft images and videos at low resolution, such as 480p, until the prompt and action work. Then render approved material at higher resolution, such as 1080p. For a lower-cost workflow, he suggests staying at 720p and upscaling later.
• Iterate deliberately. Review every generation before spending more credits. Change unclear action, simplify overlong prompts, and take a break rather than repeatedly submitting a prompt that is not working. Running multiple generations concurrently can save time when the plan allows it.
• Connect shots through action. Let movement continue across the end of one shot, then generate a complementary angle that captures the same action. This can create a match cut and smoother transitions, even if the edges of each generation contain imperfect material.
• Treat audio as a separate challenge. Joey notes that generated voices and unwanted background music can be difficult to control. He describes using Moises.ai to separate dialogue, sound effects, and music, removing the unwanted soundtrack, and then adding music or sound design during editing.
• Let people shape the final edit. Joey advises against relying on AI to edit the story. Human editors should decide the pacing, mood, meaning, and use of metaphor. He recommends organizing footage, audio, dialogue, music, sound effects, and sequences into separate folders.
• Finish with cinematic texture. The project uses film grain to soften visual inconsistencies and add texture, and black bars to create an anamorphic presentation. Joey describes placing cropped black solids over a 16:9 image, while noting that generations can also be made in a wider aspect ratio.

The overall process is:

Idea → script → beat sheet → asset tagging → reference elements → prompts → low-cost tests → refined generations → edit and sound design

Joey stresses that preparation is intended to save time and credits during generation—and that the filmmaker’s own taste and storytelling instincts remain essential.

Creators and contributors featured

• Joey — Host and creator presenting the workflow.
• KY — Collaborator involved in developing the project.
• The CTRL team — Credited collectively as collaborators on the short film and workflow.
• Higgsfield — Video sponsor.
• Robert McKee and Blake Snyder — Authors of the recommended storytelling books Story and Save the Cat!.
• Moises.ai — Audio-separation tool mentioned in the workflow.