Coze Workflow in Action: Auto-Generate 'Fat Cat at Work' Short Videos at Scale

Starting with a Viral Content Trend

Recently, the "fat cat at work" series has been absolutely exploding on short video platforms — one account grew 630,000 followers and racked up nearly 6 million views with just 280+ videos of this type. The core appeal is simple: a chubby orange tabby cat dressed in various work uniforms, "diligently working" in different settings.

This kind of "templated content" operation is a textbook example of the industrialization trend sweeping short-form video. Traditional video production requires planning, filming, editing, and music — a single video can take hours or even days. AI workflows change the game entirely: lock in a proven content structure (like "cute pet at work"), swap out variables (setting, outfit, character), and mass-produce differentiated content at near-zero marginal cost. This enables high-frequency posting and rapid account cold-starts. MCN agencies and matrix account operators have already adopted this model widely.

Today, we'll break down how to build an automated pipeline in Coze — just a dozen or so nodes — that takes a scene description as input and outputs a complete short video, end to end.

Overall Coze Workflow Architecture: A Six-Step Design

Coze is an AI application development platform launched by ByteDance. It lets users build complex AI automation pipelines through a visual node-based interface — no coding required. The core concept is the "workflow": chaining together multiple AI capabilities (LLM calls, plugin calls, conditional logic, loops, etc.) into a repeatable production line. Unlike traditional RPA (Robotic Process Automation), every node in a Coze workflow can embed AI reasoning, giving the entire pipeline semantic understanding and content generation capabilities — not just mechanical data shuffling.

This Coze workflow takes just three inputs:

• API Key: for calling the third-party image generation plugin
• Work Scene: e.g., "working at Pang Dong Lai supermarket" or "working at a dumpling shop"
• Outfit/Appearance: what kind of work uniform the cat is wearing

The workflow executes six steps in sequence: Generate image prompts → Generate storyboard images → Generate video motion prompts → Batch generate storyboard videos → Search for background music → Compose CapCut draft. Each step maps to a dedicated "worker node" with a clear division of responsibilities.

Step 1: Use an LLM to Generate Storyboard Image Prompts

The workflow starts with an LLM node (Worker #1), which takes the input scene and outfit description and generates image prompts for 5 storyboard frames.

Model choice: Doubao 1.6 Deep Thinking model. This model performs reliably when it comes to understanding complex instructions and generating structured output.

Prompt design highlights:

• Define the role as "a text-to-image prompt expert specializing in realistic fat cat work scenes"
• Instruct the model to produce 5 storyboard prompts based on scene atmosphere, actions, and the cat's outfit (adjustable to 10–15 as needed)
• Include a fixed opening description to anchor the fat orange tabby's core visual characteristics
• Provide output examples as reference to ensure consistent formatting

Key configuration: The output variable type must be set to Array<String> (array format), since multiple storyboard prompts are generated and need to be processed in a loop downstream.

Step 2: Use the Jimeng Plugin in a Loop to Generate Storyboard Images

Once the prompt array is ready, a loop node (Worker #2) processes each prompt one by one to generate the corresponding storyboard image.

There's an important plugin consideration here: a previously popular image generation plugin is being deprecated, and Doubao's Seed 3.0/4.0 costs around 260 resource points per call — relatively expensive. The recommended alternative is the Jimeng (即梦) plugin, which offers significantly better value for money.

Jimeng plugin configuration highlights:

• Reference the API Key from the start node
• Pull the image prompt from the data received by the loop
• Select "Image 4.0" as the model
• Set the aspect ratio to 9:16 (standard vertical short video format)
• Name the output variable image_list to store the generated image URLs

Step 3: AI Visual Understanding to Generate Video Motion Prompts

With static storyboard images in hand, the next step is to generate a corresponding video motion description for each image — telling the video generation model "how this scene should move." This step is the prerequisite for Image-to-Video (I2V) generation.

I2V uses a static image as the anchor for the first frame, then applies a diffusion model to predict the motion trajectory of subsequent frames, producing a coherent animated sequence. Compared to pure text-to-video generation, I2V has a key advantage: the first frame locks in the character's appearance and scene composition, dramatically reducing the chance of character "drift" (inconsistent appearance across frames). This workflow leverages exactly that — first using text-to-image to fix the orange cat's look, then using image-to-video to bring it to life, ensuring visual consistency across storyboard frames.

This step uses a batch processing node (Worker #3) with an embedded LLM node.

Model choice: Doubao 1.5 Pro Vision model. This model is chosen specifically for its visual understanding capability — it needs to "see" and interpret the image before generating a sensible motion description.

Batch processing node configuration details:

• Set parallel runs to 3 (lower concurrency to reduce errors)
• Input two data sources: text prompts from Worker #1 + images from Worker #2
• Add the image variable in the visual understanding configuration
• System prompt defines the role as "a visual script expander specializing in image-to-video prompts for fat cat scenes"

Step 4: Batch Generate Storyboard Videos

This is the most resource-intensive step in the entire Coze workflow. Worker #4 also uses a batch processing node, pairing each image with its video prompt and calling the video generation plugin.

Key video generation configuration:

• Set parallel runs to 1: generate one storyboard video at a time to avoid resource contention and errors from concurrency
• Select "Image + Text to Video" mode
• Two model options are available — Light and Pro; Pro offers higher quality but consumes more resource points
• During testing, use 480p resolution; upgrade for final production
• At 480p with the default model, each storyboard video costs under 500 resource points

Setting concurrency to 1 essentially downgrades parallel processing to serial processing — trading speed for stability. This is a common engineering tradeoff when calling third-party AI generation APIs. Higher concurrency means faster output, but it also puts more instantaneous load on the API server, making it more likely to trigger rate limiting errors that can crash the entire workflow.

The first-frame image is referenced from Worker #2's output, and the prompt is referenced from Worker #3's motion description — together, they make the static image "come alive."

Step 5: AI-Powered Background Music Matching

Worker #5 is an LLM node responsible for recommending suitable background music keywords based on the work scene. The role is defined as "an experienced music recommendation expert skilled at selecting warm, upbeat trending background music from Douyin