AI Short Video Workflow: Script, Storyboard, Generation, Editing, Publishing

A six-step pipeline that actually works

By the end of this article you'll be able to produce a 30-60 second vertical short video on a single computer, with no filming and no on-camera appearance. The pipeline breaks into six steps: script → storyboard → image generation → image-to-video → editing → publishing. The order matters more than the tools; get it wrong and every later step turns into rework.

This suits knowledge voiceover, product showcases, and scenic or atmospheric content. It is a poor fit for topics that need a real person vouching for them or precise product detail.

Step 1: write the script and storyboard with an LLM

Start by having an LLM (DeepSeek, Tencent Yuanbao, or Doubao all work) produce three things: the voiceover script, a storyboard table, and a visual description for each shot.

Four columns are enough: shot number, duration, visual content, and voiceover or subtitle. Tell the model to cut at 3-8 seconds per shot; a 30-second piece works out to roughly 5-8 shots.

You can copy this prompt directly: Turn the following content into a 45-second vertical short video script and output a storyboard table with 3-8 seconds per shot. Columns: shot number / duration / visual / voiceover. Visual descriptions must specify subject, action, environment, and lighting.

Step 2: generate images first to lock down characters and settings

This is the step most worth doing. Go straight to text-to-video and your character gets a new face in every shot; generate the keyframes as images first, then animate them with image-to-video, and the character stays consistent.

Jimeng's image tools support text-to-image and image-to-image, plus background replacement and style consistency, which makes it a good fit for producing a set of keyframes with a unified look. Keep the same appearance description for the same character and refine it through repeated passes.

Step 3: image-to-video, one shot at a time

Drop the keyframes into Jimeng or Kling and describe how the camera moves and how the light changes. Default to image-to-video; reserve text-to-video for empty scenes and atmospheric shots.

• Describe camera movement concretely: push in, pull out, pan, truck, orbit, follow. Far more useful than "cinematic."
• Describe light as a change: golden-hour side backlight, soft window light, neon reflections. Models respond to lighting more reliably than to faces.
• Change one variable at a time. If the frame is wrong, fix the description; if the motion is wrong, fix the motion. Don't move three things at once.

Step 4: assemble the fragments in CapCut

The generated clips are scattered pieces; CapCut is what turns them into a single video: stitching, trimming heads and tails, voiceover, subtitles, music, transitions, and the cover. CapCut is free and watermark-free, works on both phone and desktop, lets you start a 9:16 vertical project directly, and exports at 1080P.

Use text-to-speech for the voiceover and smart subtitles for captions; both cut post-production time significantly. Don't overdo transitions; hard cuts plus a dissolve hold up better.

Choosing tools: Jimeng, Kling, CapCut

• Jimeng: strong grasp of Chinese prompts, excellent detail in traditional Chinese aesthetics, landscapes, and food, up to 1080P, roughly 60 seconds per clip, supports first-frame plus last-frame control. Best when visual quality comes first.
• Kling: better character motion and more natural expressions, less prone to face collapse, supports motion and pose control, up to about 3 minutes. Best for content with people on screen or action demonstrations.
• CapCut: doesn't generate anything, it closes the loop. Auto movie, AI voiceover, smart subtitles, and digital human presenting all live here.

Recommended combination: Jimeng for images, Kling for character shots, CapCut for assembly. Quotas and pricing change with each version; follow the official sources (as of 2026-11).

Copyright and voice safety: don't trip here

For commercial use, stick to licensed assets provided by the platform and official voice presets. Cloning someone else's voice, reusing unlicensed music, or cutting in film and TV footage all carry infringement risk. Prefer built-in voice presets for narration and platform music libraries for soundtracks; when in doubt, switch to a stock site cleared for commercial use.

Publishing and review

Cover and title determine click-through, and the first 3 seconds determine completion. Re-publish the same video with a new cover and opening, then compare the numbers. For more tools, see the tool library; to practice the full workflow in order, start with the tutorials.