Three Prompt Templates Control All Variables for Characters, Scenes, and Image Generation
Three sets of prompt templates cover all text requirements in AI filmmaking, from shot to image to dynamic video.
They are not designed arbitrarily but correspond to three stages of the data flow: process.aiwords handles the video generation stage, places.aiwords handles the image synthesis stage, and arcpictask.taskdes handles the first-and-last-frame generation stage.
Mastering these three templates means mastering the textual control core of the entire image-generation and video-production pipeline—no other fields require manual intervention.
The most noteworthy point is that these three templates are not freely written prompts but highly structured, fixed skeletons. Every fill-in position has a clear field source. Fixed positions never change, and replaceable positions only pull values from corresponding fields in the database tables.
This design turns prompt engineering from an art into a parametric operation.
First, look at process.aiwords, the Jimeng video prompt. This is the largest in structure but has the fewest variables among the three templates.
It consists of an eight-segment skeleton, in order: character information, lip-sync consistency, accent and tone, camera technique, scene location, on-screen action, aspect ratio, clarity, and duration.
Note that this is actually nine items, but the last three are usually merged into a single technical-specification segment, hence the common name "eight segments."
Among these, character information, scene location, and on-screen action are the variables; the remaining six are all fixed.
Character information takes its value from the appearance field in dbvideo_roles—for example, an elder in a gray robe, white hair and long beard, leaning on a wooden staff.
Scene location takes its value from the environment field in dbvideo_places—for example, a stone platform on the mountainside of Cangwu, surrounded by drifting mist.
On-screen action takes its value from the actiondesc field in the process table—for example, the elder takes a banner from his sleeve.
The remaining segments are fixed: lip-sync consistency is set to enabled; accent and tone is fixed as steady Mandarin male voice; camera technique is fixed as medium shot, slow push-in; aspect ratio is fixed at 16:9; clarity is fixed at 1080p; and duration is fixed at 6 seconds.
These fixed values do not change throughout the entire book production process.
So for every shot, the prompt in process.aiwords is essentially just filling three field values into three placeholders, with all other text stitched together unchanged.
Now look at places.aiwords, the image-referencing synthesis prompt used during image-to-image synthesis.
Its most distinctive feature is the @ symbol used for image references.
BODY:
In the synthesis prompt, the @ character embeds already-generated image paths directly into the text. For example, @dbvideo_roles/fmpic_0231.png and @dbvideo_places/plcpic_087.png appearing together indicates that this synthesis uses these two images as base layers.
places.aiwords has two usage branches. When the input character field rolex is empty, it takes the pure text-to-image branch—no @ image references are present, only the location environment description, such as: at the foot of Cangwu Mountain, lush vegetation, thin morning mist, winding mountain path.
When rolex is non-empty, it takes the character-present branch, i.e., scene synthesis. In this case, the @ symbol dynamically appends the character image path corresponding to the current rolecode, along with the location image path corresponding to the placecode.
The fixed portion of the template includes a unified style instruction, such as: maintain character facial consistency, preserve location perspective, blend light and shadow naturally.
The variable portions are only the two @ reference paths, plus a single action description tied to those two paths, which is taken from the scenetopic field in the process table.
So places.aiwords does not describe the full shot. It is simply a technical synthesis prompt responsible for fusing two independent images into one, outputting a new composite image.
Finally, look at arcpictask.taskdes, the first-and-last-frame generation prompt, which controls the specific composition of the generated images.
Its key requirement is the content of the first frame and the last frame individually, as well as the difference between the two.
The fixed portion of taskdes includes output dimensions, frame sequence markers, and image style constraints.
The variable portion includes the first-frame description and the last-frame description.
The first-frame description is taken from the first-frame variant of scenetopic in the process table—for example: Lin Xuan stands before the stone platform, right hand gripping the banner, facing forward at eye level.
The last-frame description is taken from the last-frame variant of the same scene—for example: Lin Xuan turns sideways, unfurls the banner, gazing toward the cliff in the distance.
taskdes must explicitly label which description corresponds to the -a output and which to the -b output; otherwise, the system cannot automatically generate the two files.
Additionally, taskdes includes an immutable prohibition: the generated images must preserve the character's identity and the location's layout exactly as in the reference images—only pose and camera angle are allowed to vary.
This constraint ensures that the first and last frames maintain visual unity while exhibiting motion-based differences.
Comparing the three templates side by side reveals their respective layers.
places.aiwords is the underlying image-to-image tool—it directly composites the character image and location image into a single frame featuring the character within the scene.
BODY:
arcpictask.taskdes is the mid-level control layer. It builds on the output of places.aiwords to further refine the specific composition requirements for the first and last frames, ensuring the two images form a pair of frames linked by meaningful change.
process.aiwords is the top-level video layer. It breathes life into the shot, telling the Jimeng video generator how to fill in the motion between the first and last frames.
The hierarchical relationship of the three templates also mirrors the production order: first run places.aiwords to generate the composite image, then run arcpictask to generate the first and last frames, and finally run process to generate the video.
Each layer receives the previous layer's output as its input while carrying its own fixed format framework.
From the field-source perspective, all variable values come from fields in the dbvideo_process table.
The three variables in process.aiwords correspond to role_name, place_name, and action_desc, respectively.
The two @ references in places.aiwords correspond to fmpic and plcpic, along with scenetopic.
The variables in arcpictask.taskdes correspond to first_frame_desc and second_frame_desc. These fields are already populated when atctoscene generates the storyboard, or are automatically extracted from the original source content by auxiliary scripts in subsequent steps.
Manual intervention is extremely low. Only when initially establishing the templates must fixed text be written; afterward, the prompts for all 1,029 shots are completed entirely through field concatenation.
This fundamentally changes the traditional AI-image-generation approach of writing a single prompt as one long descriptive paragraph, instead breaking it down into multiple reusable fixed skeletons plus a few variables.
The benefit is greatly improved stability—no shot's prompt will be so freely written that it causes a split in art style.
When changing a character, you only need to modify the appearance field in dbvideo_roles; all process.aiwords entries referencing that character will update automatically, with no need to rewrite any prompt.
The field structures of the three templates can be summarized in a simple table: process.aiwords—fixed eight segments, with replaceable variables of character information, scene location, and on-screen action.
places.aiwords—fixed style sentences, with replaceable variables of two @ reference paths and one action sentence.
arcpictask.taskdes—fixed frame dimensions and style instructions, with replaceable variables of first-frame description and last-frame description.
BODY:
Each template has no more than three variables, compressing the prompt complexity of the entire project to an extremely low level.
This means that even someone with no prior experience in AI prompting can independently generate complete image-generation and video prompts for all 1,029 shots, as long as they understand the field sources and filling logic of these three templates.
Prompts are no longer artistic expression—they become data-driven, parameterized configuration.
This is the fundamental reason the entire system can be run by a single person on three computers: its control surface is extremely limited.
Mastering these three templates turns the production of every shot into a mechanical routine of template filling and interface calls.