“For we wrestle not against flesh and blood, but against principalities, against powers, against the rulers of the darkness of this world…” — Ephesians 6:12
Stand Firm is a short AI film about the unseen war: an army of darkness waiting in the fog, and the armies of heaven tearing through the sky to meet it. I didn't write a single model prompt by hand. I directed it the way a director works with a crew: I described the story, approved or rejected every image, and gave notes. My AI agent wrote the prompts, attached the reference images, and ran the generations. This guide gives you that system: an image-to-video workflow that works for any script, plus copy-and-paste prompts to give your agent at every step, and the prompts I gave mine.
Before You Begin…
- Platform: Higgsfield
- Workflow: Image-to-video
- AI Agent: Claude, connected to Higgsfield so it can write prompts, run generations, and check results for you
- Image Model: Seedream 5.0 Pro, 2K, 16:9
- Video Model: Seedance 2.5, 1080p, 16:9
- What You Bring: A script (even a one-paragraph story works) and one reference image that captures the look you want
- Next Step: Stitch the finished videos in your editor, add your music or voiceover, and post.
The whole method comes down to one chain: script → segments → images → videos. Every segment of your script gets one image, and every image becomes the start frame of one image-to-video generation. You direct the images first, then animate them. Images are fast and cheap to fix. Video is not.
Step 1: Brief Your Agent
Start by setting up the collaboration the way you would brief a crew: the story, the look, the characters, the models, and the rules that apply to every generation. Ask the agent to describe what it sees in your reference images and confirm before it does anything. That first reply tells you whether it actually understood your world.
Agent prompt template
I'm making a short cinematic [TRAILER / SHORT FILM / MUSIC VIDEO] about [YOUR STORY IN ONE SENTENCE]. I'll share my reference images and my script. I need your help generating an image for each scene and then animating each one in Higgsfield with image-to-video. Look: [ATTACH YOUR ANCHOR IMAGE AND SAY WHAT MATTERS: color, lighting, lens, mood] Characters: [ATTACH YOUR CHARACTER SHEETS OR LIST THEIR HIGGSFIELD IMAGE IDS] Image model: [MODEL, QUALITY, SIZE] Video model: [MODEL, RESOLUTION, ASPECT RATIO] Rules for every generation: [e.g. no music, no text, no sparks] Tell me what you see in my reference images and confirm you understand before we start.
What I told my agent
Here's a battle scene I made between two forces: light versus dark, good versus evil. This is the look for the whole project. I'm making a music video, and I need your help generating an image for each scene and then animating them. I already have character sheets for both sides. Tell me what you see before we start.
My anchor was one battle image and its mirror: the exact same composition generated twice, with only the armor swapped. If your story has two sides, do the same. Everything lines up when you cut, blend, or split between them later.
Step 2: Lock Your Reference Images
Give your agent a character sheet for every character or faction: full body, close-up, and one more angle. These become the design reference for every image you generate. If you made them in Higgsfield, share their image IDs so the agent can attach them directly.
Agent prompt template
Here are my reference images: [ATTACH OR LIST IMAGE IDS, ONE PER CHARACTER OR FACTION, PLUS YOUR ANCHOR IMAGE]. Use them as the design reference for every image so the characters never drift. When we move to video, use single images as references instead of the multi-panel sheets.
What I told my agent
Here are my reference images, all generated in Higgsfield, with their image IDs: - Dark character sheet: [image ID] - Light character sheet: [image ID] - Dark battle scene: [image ID] - Light battle scene: [image ID] Use these for every image so the characters and the look stay consistent.
Step 3: Break Your Script Into Segments
This is the heart of the system. A segment is one continuous moment of your story, usually 4 to 15 seconds long. Each one needs a clear action, one image, and a camera feel. Describe your story in plain words and let the agent turn it into a segment plan, but don't let it generate anything until you approve the plan.
Agent prompt template
Here is my script: [PASTE YOUR SCRIPT OR DESCRIBE THE STORY IN PLAIN WORDS] Break it into segments of [4–15] seconds. For each segment, give me: - what happens - the image we need (shot size, angle, subject) - the camera feel (slow and steady for tension, or handheld and chaotic for action) - the duration Don't generate anything yet. I want to approve the plan first.
What I told my agent
Here's the story. The dark knights stand in battle formation in a dark, foggy, open space, with torches held up along the front line. Everything is still and orderly. Then a blast rips through the sky, and in that one blast of light, an army of white angelic knights appears in front of the battlefield, all at once, not one by one. Ultra-wide shot as both front lines face each other. Once the battle begins, it's epic and chaotic: cinematic, low-angle, Dutch-angle shots. Break this into segments before we generate anything.
That paragraph became this segment plan:
| Segment | What happens | Camera feel |
|---|---|---|
| 1. The dark army waits | Ranks in formation, torches on the front line, fog everywhere | Slow creep, tension |
| 2. The sky tears open | A blast of light rips through the storm clouds | Controlled, awe |
| 3. The light army arrives | One blast of light populates the white army in formation | Controlled, awe |
| 4. The standoff | Both front lines face each other across the mud | Slow push, tension |
| 5. The clash | The lines collide | Handheld chaos |
| 6. Victory | The white army stands over the fallen dark army | Slow and steady |
Working to music? Share the song's timing with your agent so segments land near the key moments, then fine-tune the timing yourself in the edit. That's what I did.
Step 4: Generate One Image per Segment
Now the agent turns each segment into an image. These images become the start frames for your videos, so this is where you lock the look. Ask for all of them at once so you can judge them side by side, and treat it as a test of whether the agent has the idea locked in.
Agent prompt template
Generate one image for each segment using [MODEL] at [QUALITY], [ASPECT RATIO]. Use my reference images so the look stays consistent. I need [NUMBER] images to see if you have the idea locked in. Show me all of them together so I can approve or reject each one.
What I told my agent
Generate five images, one per segment, so I can see if you have the idea locked in. Use Seedream 5.0 Pro at 2K, 16:9 (2720×1536), with my character sheets and battle scenes as reference images.
Step 5: Give Notes Like a Director
You don't need to know prompt syntax to fix an image. Say what's wrong the way you'd say it on set: what to remove, what to add, how the composition should feel. The agent translates your note into the prompt.
Agent prompt template
All good except for [IMAGE]. I do not want [WHAT TO REMOVE]. Re-render that image. Also, [COMPOSITION NOTE: e.g. add out-of-focus foreground objects for depth instead of an evenly framed image].
What I told my agent
Everything works except the clash. Remove the sparks and re-render that image. Also, the composition feels flat and evenly framed. Add out-of-focus objects in the foreground to create depth.
Step 6: Turn Each Image Into Video
Each approved image becomes the start frame of its segment's video. Describe the feel of the camera, not the prompt wording: where tension builds, where chaos breaks loose, and what should never appear. Set those rules once and ask the agent to apply them to every generation.
Agent prompt template
Animate segment [NUMBER] using its image as the start frame on [VIDEO MODEL] at [RESOLUTION], [DURATION] seconds, with [NUMBER] hard cuts inside the generation so it's a ready-to-go edit. During [ESTABLISHING MOMENTS], I want tension to build: slow, steady camera. During [ACTION MOMENTS], I want [e.g. chaotic, unstable camera, fast pans, drifts, speed ramps]. Add [THINGS TO AVOID] as negative prompts in every generation.
What I told my agent
Animate each image on Seedance 2.5 at 1080p, using it as the start frame. Rules for every video: add "no music" as a negative prompt, and no sparks during the battle. In the establishing scenes, build tension with a slow, steady camera. When the battle starts, all hell breaks loose: chaotic, unstable camera, fast pans and drifts.
Step 7: Fix What Comes Back Wrong
Some videos will come back looking like a still picture with almost no motion. Don't start over. Ask for a makeup generation from the same image and list the small movements you expect to see. That list is what gets the shot moving.
Agent prompt template
Segment [NUMBER] came out as a still picture, not moving. Regenerate it from the same image with minor movements the whole time: [LIST THEM: capes in the wind, fog drifting, flags waving, horses shifting]. Then hard cut to [NEXT MOMENT] with [CAMERA FEEL]. [NUMBER] hard cuts in total, [DURATION]-second generation.
What I told my agent
Some videos came out as still pictures, so let's do makeup generations. For the standoff, start from that image with a smooth push forward down the lineup. Keep minor movements going the whole time: capes moving in the wind, archers drawing their bows, flags waving, horses shifting, fog drifting under their feet. Then hard cut to the middle of the battle, like the reference image attached, with speed ramps, whip pans, and intense, violent camera motion. Four hard cuts in total, 11 seconds.
Step 8: Add Inserts
Inserts are the short detail shots that give an edit texture. Ask for a few image options, pick the strongest, then animate it as one continuous shot.
Agent prompt template
Generate an image option that shows a close-up of [DETAIL] in the foreground, in focus, with [BACKGROUND] out of focus. I need it as an insert shot. Then give me [NUMBER] more options from the [SCENE]: close-up and macro close-up shots, [ASPECT RATIO]. I will pick the best image to animate.
Give me a [DURATION]-second animation of the [SHOT]: handheld, close up on [SUBJECT], shallow depth of field. One continuous generation, no cut. At [TIME], [ACTION].
What I told my agent
Generate an insert image: a close-up of one torch in the foreground, in focus, with the army out of focus behind it. Then give me two more options from the standoff scene where both armies face each other: one close-up and one macro close-up, 16:9. I'll pick the best image to animate.
Give me a 4-second animation of the torch image: handheld, close up on the torch, shallow depth of field. One continuous shot, no cut.
Now animate the helmet close-up for 4 seconds, one continuous shot, focused on the knight. At 0:01, he lifts his sword and yells, starting the clash.
Step 9: Land the Ending and the Title
Close the story with its own video, then give it a title card in the same look as the film.
Agent prompt template
For the last one, a [DURATION]-second generation with [NUMBER] different [ENDING] shots: [DESCRIBE THE FINAL STATE OF YOUR STORY].
Generate a [DURATION]-second opening title that says "[TITLE]" in the same look as the film.
What I told my agent
For the final video, generate 12 seconds with four different victory shots: the white army standing victorious over the dark army, who have all fallen on the ground. No music.
Generate a 3-second opening title that says "Stand Firm" in the same look as the film.
Step 10: Edit and Package
Stitch the videos in order, add your score or voiceover, and export. Then keep the agent working for you: titles, captions, and social assets built from the same images.
Agent prompt template
Give me [NUMBER] title options for this short film.
Give me [NUMBER] caption options for Instagram. My audience is [WHO THEY ARE] and my goal is [WHAT YOU WANT THE POST TO DO].
Build me a [NUMBER]-tier [ASPECT RATIO] collage with no borders from the still images. Give me as many variations as you can.
What I told my agent
Give me four caption options for Instagram. I make short AI films to get the attention of studios looking for new ways to create visuals with generative AI.
Build a three-tier 9:16 collage with no borders from the still images. Give me as many variations as you can.
What I Learned
- Direct the images first. Fixing an image takes one quick render. Fixing a video takes a full generation.
- Give notes, not prompts. Tell your agent what you want to see and feel, and let it handle the model language.
- Mirror your anchor image. Two sides of a story in one shared composition gives you cuts, blends, and splits that line up perfectly.
- Name what moves. If a video comes back frozen, list the small movements you expect to see and regenerate.
- Set your rules once. "No music" and "no sparks" went into every generation because I said it once, up front.
- Split your camera language. Slow and steady for tension, handheld and chaotic for action. The contrast is what makes the story hit.
Build Your Own
Start with one image you love and a story you can tell in one paragraph. Break it into segments, direct the images, and turn each one into video inside Higgsfield AI. Your agent handles the prompt language. You handle the vision.
Affiliate disclosure: If you sign up through my link I may earn a commission at no extra cost to you. It helps support more free AI filmmaking resources like this one.