How to Use an LLM Like Claude to Write Your Story: Generate Prompts Scene by Scene

Most people think of AI video tools as the star of the show. But there's a smarter way to work: use a large language model like Claude as your creative director and prompt factory, and let the video tools do what they're actually good at — rendering.

The result is something genuinely new: a structured, consistent AI-generated story where every scene feels connected, every shot is purposeful, and the video prompts you feed into Sora, Kling, or Runway actually hold together as a narrative.

Here's how to do it.

Why You Need an LLM in Your Video Workflow

When most creators try to make an AI video story, they jump straight into the video tool and start prompting scene by scene. The problem is that each prompt exists in isolation. The AI has no memory of the last scene. The lighting shifts. The character's coat changes color. The tone drifts from tense to cheerful for no reason. By scene five, you've lost the thread entirely.

An LLM solves this by acting as the connective tissue of your story. Before a single video gets generated, Claude holds the whole narrative in its context — the characters, the world, the emotional arc, the visual language — and uses that complete picture to write every video prompt with consistency baked in.

Think of it less like chatting with an AI and more like briefing a screenwriter who also happens to know exactly how to talk to a video model.

Step 1: Build Your Story Bible Inside the Conversation

Before you ask Claude to write a single prompt, you need to establish the world. This is your story bible, and you build it through conversation.

Start by telling Claude the big picture. What is this story about? What genre is it? What's the emotional journey? Then go deeper:

Characters: Give each character a detailed physical description that will stay locked across every scene. Don't just say "a young woman." Say: *"Maya, late 20s, short natural hair, always wearing a worn olive green jacket, has a small scar above her left eyebrow, moves with quiet intensity."* The more specific you are here, the more consistently Claude will describe her in every prompt it writes.

The world: Establish the visual rules of your setting. Is it a near-future city with neon and rain? A sun-bleached desert town? A cozy but slightly off-kilter suburb? Give it texture — the colors that dominate, the time of day, the atmosphere.

Visual style: Define a consistent cinematic language. Are you going for gritty handheld realism? Wide, composed shots with still cameras? Warm golden-hour lighting throughout? These details become part of every prompt Claude generates.

Tone and pacing: Is this a slow burn thriller? A melancholy character study? An action-driven chase? The emotional register affects how Claude frames each scene — the shot distance, the urgency of the language, the sensory details it includes.

Once you've laid all of this out, ask Claude to confirm it back to you in a structured summary. This becomes the shared reference point for everything that follows.

Step 2: Break Your Story Into Scenes

Now work with Claude to map out your story structure. You don't need a full screenplay — you need a scene list. Think of it as a shot list for a short film or episode.

Ask Claude to help you break the story into discrete moments, each one a self-contained event that can become a video clip. A good scene for AI video generation is usually 5–15 seconds of focused action: a character entering a room, a confrontation at a doorway, a quiet moment by a window, a chase through a narrow alley.

For each scene, establish:

• What is happening physically (the action)
• Where it is set (the specific location)
• What the emotional function of the scene is (what feeling it creates, what it advances in the story)
• How it connects to the scene before and after it

This scene map becomes Claude's working document. Every prompt it writes will reference back to it, ensuring that scene seven doesn't accidentally contradict scene three.

Step 3: Generate Your Video Prompts Scene by Scene

This is where the workflow pays off. With your story bible and scene map established, ask Claude to write a video generation prompt for each scene.

A well-structured video prompt contains several layers that Claude can now write with full context:

The action: What is happening in the shot, described with visual specificity. Not "Maya looks worried" but "Maya stands at the rain-streaked window, arms crossed, jaw tight, eyes tracking something unseen in the street below."

The shot framing: The camera perspective and distance. Medium close-up. Wide establishing shot. Low angle looking up. Over-the-shoulder. This creates cinematic intentionality across your sequence.

The lighting and atmosphere: Consistent with your visual bible. If your story is set in golden-hour warmth, every prompt reinforces that. If it's cold blue urban night, the prompts say so every time.

The character description: Even though you've established your characters, Claude should include their key visual anchors in every prompt — the olive jacket, the scar, the natural hair — because the video model has no memory of previous generations. These details are the consistency mechanism.

The environment: Specific details of the location that match your world-building.

A finished prompt from Claude might look like this:

"Cinematic medium shot, slightly low angle. Maya (late 20s, short natural hair, worn olive green jacket, small scar above left eyebrow) stands frozen in the doorway of a dimly lit concrete stairwell. She grips the metal railing. Behind her, emergency lighting casts a red glow. Her breath is visible in the cold air. She is listening for something. Tense, still, hyperaware. Gritty urban realism, desaturated color grade with deep shadows."

That prompt can go directly into Sora, Kling, or Runway. It contains everything the video model needs and nothing extraneous.

Step 4: Maintain Continuity Across the Sequence

As you generate prompts for later scenes, Claude can track what has already happened in the story and adjust accordingly. A character who was injured in scene four should still show signs of that in scene seven. A location visited earlier should have consistent details when revisited. Time of day should advance logically.

This is where the LLM genuinely earns its place in the workflow. Because Claude holds the entire conversation in context, you can simply say *"write the prompt for scene eight, where Maya returns to the stairwell but this time she's confident rather than afraid"* — and Claude will produce a prompt that echoes the earlier scene's visual details while shifting the emotional register.

You can also ask Claude to flag potential continuity issues. *"Are there any contradictions between the prompts so far?"* or *"Does the lighting stay consistent across the outdoor scenes?"* These are exactly the kinds of editorial questions an LLM can help you answer before you spend time generating video.

Step 5: Export Your Prompt Package

Once all your scene prompts are written, ask Claude to compile them into a clean, organized document — scene number, brief scene description, and the full video prompt, all in order. This becomes your production document: the thing you work from as you move into the video generation phase.

Having everything in one place means you can generate scenes in any order, come back to the project days later, or hand it off to someone else, without losing coherence.

Tips for Getting the Best Prompts from Claude

Be a demanding collaborator. If a prompt Claude writes feels generic or loses a character detail, push back. *"This prompt doesn't mention the scar. Add it. Also make the framing tighter — I want a close-up, not a medium shot."* Claude responds well to specific, direct feedback.

Ask for variations. For key scenes, ask Claude to write two or three different prompt versions with different shot choices. *"Give me the same scene from a wide establishing shot and also from a tight close-up on her hands."* This gives you options when you're in the generation phase.

Use Claude to reverse-engineer prompts from references. If you have a film still or a screenshot that captures the vibe you're after, describe it to Claude in detail and ask it to write a prompt that captures that visual language. *"I want the same feeling as a moonlit scene with long shadows and a single light source — write a prompt that builds that atmosphere into the scene."*

Lock your character descriptions as a system. At the start of your conversation, paste your full character descriptions and tell Claude: *"These descriptions are locked. Include the relevant anchors in every single prompt you write, regardless of scene."* This creates a standing instruction that Claude will honor throughout the conversation.

Think in transitions. Ask Claude to consider how scenes will cut together. *"The previous scene ends with Maya running. This scene should open with her already stopped, catching her breath — write the prompt to establish that transition."* This kind of editorial thinking, built into the prompting phase, makes the final edit far smoother.

The Bigger Picture

What this workflow gives you is something that hasn't really existed before: a structured creative process for AI video storytelling that treats the LLM as a director and the video model as a camera.

Claude doesn't just write prompts — it holds your story's logic, remembers your characters, enforces your visual language, and helps you think through narrative decisions before you've committed to any generated footage. That's not a helper. That's a collaborator.

The video tools are getting better every month. Sora 2, Kling 2.6, Runway Gen-4 — they can all render scenes of remarkable quality. But quality footage without narrative structure is just a highlight reel. The story — the thread that makes a viewer lean in and keep watching — comes from the work you do before you ever hit generate.

And that work belongs in an LLM conversation.

Start with your story bible. Build your scene map. Let Claude write your prompts. Then generate. That's the workflow — and it's more powerful than any single AI video tool on its own.