Archival Fragments (Simplified Version)
This skill locks in only two things: this visual style and the reference images → start/end frames method.
There is no model lock-in, no dependency installation, and it does not run the workflow for you. You can use any image model or video tool, as long as the video tool supports start and end frames.
Three Hard Constraints
Break any one of them, and the finished video will be ruined. These three rules were learned the hard way from real failures.
1. Every Shot's Reference Image Must Be Different
N shots require N+1 distinct reference images.
“Two adjacent shots sharing the same image” means: Shot 1's end frame = Shot 2's start frame. That one image is shared by the two shots.
It does not mean using one image for the entire video.
Absolutely forbidden:
- Using the same image as the input for every shot
- Using a single structural diagram, schematic, or infographic as the reference for every shot
- Using
assets/style-reference.png
This actually happened: an Agent used a single structural diagram as the reference for every shot, resulting in dozens of shots that all looked exactly the same. The entire video was effectively a static image with a voice-over.
Self-check: number of reference images == number of shots + 1; no two image descriptions are duplicates; each image has its own independent prompt.
2. Prefer Concrete Subjects Over Abstract Graphics
The paper and red lines are the ground of the image, not the subject. The things placed on top must be concrete enough for viewers to understand at a glance.
Find a universal symbol for each line of copy, then converge on 2–3 symbols for the entire video, using each several times:
| Meaning | Symbol |
|---|
| Who or where, market share, from nothing to something | World map + push pins + labels |
| Price, cost, imbalance | Balance scale (one end holds physical objects, while the other is continuously loaded) |
| Ranking, progression, how much is left | Steps / podium |
| Quantity, density, sheer scale | A large pile of physical objects (chips, sheets of paper, a crowd) |
| Time point, formality | Flip board, stamp, official document |
| Denial, erasure | A hand tearing something away / white paper covering it |
Using blank paper and a line to mean “this position is empty,” or using three squares to mean “three companies”—these kinds of abstract graphics are incomprehensible to viewers and failed in testing.
3. Slow Does Not Mean Static
Every shot's final frame must be clearly different from its opening frame. Even a negative-space shot needs an action moving across the frame.
Start/end frames naturally enforce this: start frame != end frame, so the image must change.
Three-Step Workflow
Step 1: Break Down the Shots and Map the Reference-Image Chain
Break the script into shots by sentence (2–3 shots per sentence), and design a continuous visual journey:
A01 ──W01──> A02 ──W02──> A03 ──W03──> A04 ...
The end state of one shot becomes the start state of the next.
There are only four legal forms of visual progression. Every shot must fit one of them:
| Method | Example |
|---|
| Object change | Zero push pins on a map → one → three |
| Framing change | Wide shot of a world map → push in on Korea → close-up of one pinhole |
| State change | A few coins on a scale → coins stacked into a tall tower |
| Subject replacement | Map exits the frame → balance scale rises into frame |
Show the user the breakdown first—revising a shot list takes 5 minutes, while regenerating it takes several hours.
Step 2: Generate the Reference Images
Use text-to-image, with one prompt per image and no reference image attached. Each prompt = fixed style prefix + the content of that image.
See keyframes.md for the format; the style prefix is in that document.
After generation, review each image one by one: whether the text has broken, whether viewers will understand the meaning, and whether it connects cleanly to the images before and after it. Images are cheap, video is expensive, so this step cannot be skipped.
Step 3: Generate Each Shot with Start and End Frames
Attach two images to each shot: image N as the start frame and image N+1 as the end frame. The prompt should describe only the motion, in one sentence.
See shots.md for the format.
Tool Requirements
| Stage | Requirement |
|---|
| Image | Can produce the required aspect ratio (4:3 recommended) and renders text accurately |
| Video | Must support start and end frames—also called “first & last frame,” “start/end frame,” or “frames to video” |
Pure text-to-video tools that do not support start/end frames cannot execute this method; transitions between shots will inevitably jump.
Generation duration: most tools have a 5-second minimum, while each shot in this style usually uses only 1–3 seconds. Generate everything at 5 seconds, then adjust the speed in editing to the target duration—the rhythm is set on the editing timeline, not in the prompt.
Documentation
| Document | What It Covers |
|---|
| style.md | Visual baseline: color palette, the four roles of red, technical annotation vocabulary |
| keyframes.md | Reference-image chain design, style prefix, image prompt structure, text rules |
| shots.md | Start/end-frame motion sentence patterns, verb list |
is the configuration for a completed 67.3-second film: the original prompts for 35 reference images and 34 motion sentences.
Follow its structure, but do not copy its content—those symbols were designed for a specific subject, so you need to find new ones when changing the subject.
assets/example-01..08-*.jpg
are eight consecutive reference images. Look at how each one advances the sequence, and you will understand what “the image for every shot must be different” means.