The Secret of Only Three Variables in the Eight-Segment Skeleton Prompt of aiwords

In the Yuncao Academy system, the aiwords field in the dbvideo_process table does not store an arbitrary video description, but rather a Jimeng video prompt string that strictly follows an eight-segment skeleton structure. The order of these eight segments is fixed, and each segment is separated by a Chinese semicolon or a line break. After reading this field, the system submits it verbatim to the Jimeng video generation interface. The eight-segment skeleton, in order, is: character information, lip-sync consistency, accent and tone, camera technique, scene and location, on-screen action, aspect ratio, resolution, and duration. The first segment, character information, describes the core character's appearance, clothing details, identity, and temperament in the shot; this is the static definition of the character's image. The second segment, lip-sync consistency, controls the accuracy of lip movements when the person in the video speaks, typically written as "natural lip movement" or "lip-sync aligned," working with the system's built-in audio features to match speech to lip movement. The third segment, accent and tone, describes the character's vocal qualities, such as a deep male voice, a clear young voice, or a slightly hoarse elderly voice, as well as the pace of speech. The fourth segment, camera technique, specifies camera movement, including push in, pull back, tracking shot, fixed position, overhead shot, low-angle shot, and orbit; Jimeng generates corresponding camera movement effects based on these words. The fifth segment, scene and location, describes the environmental background of the current shot, precise to the name of the city or location features, ensuring the video background matches the first frame image. The sixth segment, on-screen action, describes the character's behavior in the shot, which can be verb phrases with clear direction, such as walking, turning around, laughing, waving, or drawing a sword. The seventh segment, aspect ratio, sets the picture ratio, typically written as 16:9 or 21:9. The eighth segment, resolution, and the ninth segment, duration, appear together in the final parameter description; the system automatically fills them in as resolution requirements such as 1080p or 2K, and duration such as 5 seconds or 10 seconds. Within this eight-segment structure, there are truly only three variables that the operator fills in through data each time a storyboard is split. The first variable is character information. This value is not written directly into the aiwords field, but is linked through the rolex field to a character record in the dbvideo_roles table. When the system generates the aiwords string, it extracts from dbvideo_process the character name and character description corresponding to rolex, for example, "Lin Xuan, young wandering swordsman, black hair tied in a topknot, wearing dark martial robes, long sword at the waist, cold expression," and then fills this text into the character information position. The second variable is scene and location. It is linked through the placex field to the place_desc in the dbvideo_places table, for example, "Five Elements City East Gate, built from gray boulders, wall three zhang high, copper door knockers on the gate, two stone lions before the gate, ground paved with blue stone slabs." The third variable is on-screen action. It is extracted from the original text's verb phrases during the atctoscene split, and is in essence the core narrative content of the shot. These three variables are placed into their corresponding skeleton positions, while all other segments, including lip-sync consistency, accent and tone, camera technique, aspect ratio, resolution, and duration, are entirely generated statically by the system based on imgstyle and default configuration. The elegance of this design is that it nearly eliminates human spelling errors and style drift. Suppose template-based generation were not adopted and the operator was instead required to manually enter a one-sentence prompt for each storyboard. Then 1,029 shots would produce 1,029 different expression styles. Some shots might read "the character stands at the door," others might read "Lin Xuan stands at the door," and still others might read "there is a person at the city gate," and these descriptions would produce entirely different visual results. What is more troublesome is that manual entry makes it easy to omit key information. For example, if the camera movement is forgotten, the videos generated by Jimeng tend to choose camera movements randomly, causing the continuity between adjacent shots to break. Template-based aiwords, by fixing the eight-segment structure, guarantees that every shot contains the complete parameters needed to generate video. At the same time, because the remaining six segments other than the three variables are constant, the model does not generate extra randomness due to abrupt changes in prompt structure, making the generation results more controllable. An actual concatenation process can intuitively demonstrate how this template operates. Suppose a dbvideo_process record has scenecode SCN000245, rolex pointing to the character Lin Xuan, and placex pointing to the Five Elements City East Gate. When the system reads the character table, it obtains the description: "Lin Xuan is an eighteen-year-old young man with a handsome face, sword-shaped eyebrows and starry eyes, wearing black leather armor, a silver belt at the waist, and black leather boots on his feet." When reading the location table, it obtains: "Five Elements City East Gate is a giant stone gate with carved dragon pillars on both sides; the gate is about five zhang high, with prominent door studs, an overall blue-gray tone, irregular stone slabs on the ground, and distant mountain peaks visible in the background." The on-screen action is extracted from the original text as "Lin Xuan raises his head and gazes at the city gate plaque." At this time, the imgstyle value is gz. The aiwords is concatenated according to the eight-segment template into the following complete text: "Character information: Lin Xuan, eighteen-year-old young man, sword-shaped eyebrows and starry eyes, wearing black leather armor, silver belt at the waist, period costume styling; lip-sync consistency: natural lip movement, synchronized with audio; accent and tone: low male voice, medium speaking pace, steady breath; camera technique: full-shot fixed position, main camera slowly pushing forward, slight low angle; scene and location: Five Elements City East Gate, giant stone gate, carved dragon pillars on both sides, blue-gray tone, distant mountains in the background; on-screen action: Lin Xuan raises his head and gazes at the city gate plaque, solemn expression; aspect ratio: 16:9; resolution: 2K; duration: 5 seconds." This aiwords string is then written into the dbvideo_process record and passed to the Jimeng interface. If the same shot needs to be converted to anime style, the system does not modify the aiwords field itself, but indirectly adjusts the description words in two segments, character information and scene and location, through the influence of imgstyle. Specifically, when imgstyle is dm, the character description gains anime-stylized rendering requirements, such as bright facial lines, highlights in the pupils, sharp edges on the costume, and the scene gains atmospheric light effects and a simplified, lively background. But these refinement details are all encapsulated in the data generation process of the character table and the location table. The template itself does not perceive style; it is only responsible for building the framework, and the style details come from the source data. This means the template is a static mechanism and the data is dynamic content. Only by decoupling the two does batch control become possible. If every storyboard were written from scratch as a complete prompt, the workload for 1,029 shots would be enormous. Even if each shot took only three minutes to write a prompt, it would require fifty-one hours of continuous work. Even more critical is that descriptions of the same character across different shots could easily differ from one another: one shot might read "black leather armor," another might read "dark martial robes," and yet another might read "dark-blue waist sash." These subtle differences might go unnoticed in static images, but in video, when the character's costume color and hairstyle flicker and change across different shots, the sense of realism of the entire finished film collapses instantly. Template-based prompting fundamentally solves this problem, because character information is always pulled from the same row in the dbvideo_roles table. All storyboards referencing that character receive exactly the same character description. No matter how many times generation is run, the consistency of this description will not degrade due to operator fatigue. Another point to note is that although there are only three variables in the process.aiwords field, in actual operation you will discover one hidden variable as well: the style descriptor words injected by imgstyle. These style descriptor words are generally concatenated onto the end of the character information segment and the end of the scene and location segment. For example, in period style, the scene and location is followed by appended words such as "ink wash landscape texture" and "negative space composition." In essence, this is not a fourth variable, but rather a static modifier automatically injected by the system based on imgstyle, and it does not increase the operator's workload. Therefore, the maintenance cost of the entire system is extremely low. Changing a single action in a single storyboard requires only updating the single row in dbvideo_process. Changing the visual image of an entire character requires only updating one row in the character table. Changing the visual style of the entire film requires only a batch update to the imgstyle field. All updates can be completed in seconds and take effect immediately in the next round of video generation. This design does not pursue the literary quality of the prompt itself, but rather the stability of the structure and the traceability of the data. It is precisely this engineering-oriented philosophy that makes completing a two-hour film in eleven days a reproducible operational process.