mirror of
https://github.com/magnus919/agent-skills.git
synced 2026-09-11 19:47:12 +03:00
74 lines
3.4 KiB
Markdown
74 lines
3.4 KiB
Markdown
# Video Prompting Reference
|
|
|
|
## Shot template
|
|
|
|
```text
|
|
SHOT: [duration and framing]. [Subject] remains recognizable and grounded in the
|
|
space. One [screen/projector/industrial practical] is the dominant light source.
|
|
|
|
CAMERA: slow [push-in/pull-back/lateral track], stable mechanical movement, slight
|
|
servo hesitation, shallow focus breathing, no handheld energy.
|
|
|
|
ACTION: [one restrained physical action]. Allow long pauses and incomplete
|
|
information. Reveal [specific detail] gradually, then return to an unresolved state.
|
|
|
|
SURFACE AND SIGNAL: near-monochrome charcoal, smoked gray, dirty silver, and aged
|
|
ivory; crushed blacks; clipped highlights; restrained scanlines, interlacing, grain,
|
|
dust, scratches, halation, phosphor bloom, mild jitter, and occasional horizontal
|
|
interference embedded in the recording.
|
|
|
|
SOUND: [ambient machine hum / projector chatter / room tone]. No music unless
|
|
requested. No dialogue unless a visible speaker is specified.
|
|
|
|
AVOID: frenetic cuts, whip pans, action-camera movement, constant stimulation,
|
|
bright neon, decorative digital glitch, gore, jump scares, polished CGI.
|
|
```
|
|
|
|
## Motion rules
|
|
|
|
Use one primary motion and one or two secondary changes. Primary motion can be a
|
|
camera move, mechanical actuation, gradual reveal, or a slow change in focus. Secondary
|
|
changes can be screen flicker, exposure breathing, dust crossing the beam, or tracking
|
|
instability. Do not animate every object. Tension comes from duration, darkness,
|
|
partial evidence, and the anticipation of recognition.
|
|
|
|
A loop should not resolve the mystery. End with the subject returning to its initial
|
|
configuration, the light failing to reveal the key detail, or the observer stopping
|
|
just before recognition.
|
|
|
|
## Mode selection
|
|
|
|
- **Text to video:** use when the shot is conceptual and no opening composition must
|
|
be preserved.
|
|
- **Image to video:** use when a supplied or generated still establishes the subject,
|
|
framing, light, and surface treatment. Describe only the motion and temporal change
|
|
that should occur.
|
|
- **Keyframes:** use when a transformation or state change must remain coherent. Keep
|
|
the same subject location, dominant light, and negative-space structure between
|
|
keyframes; vary only the intended state.
|
|
- **Continuation:** use when extending a shot. Preserve the established camera
|
|
behavior, exposure, sound bed, and unresolved tension rather than introducing a new
|
|
visual language.
|
|
|
|
For FLUX 3, read the `video-generation` skill and its prompting guide before invoking
|
|
any generation tool. Respect that skill's duration, frame-index, audio, and result
|
|
polling requirements.
|
|
|
|
## Video verification
|
|
|
|
Use FFmpeg or FFprobe to record container, stream, frame rate, duration, dimensions,
|
|
and audio details. Extract a bounded set of frames from the beginning, middle, motion
|
|
change, and end. Inspect those frames for:
|
|
|
|
- the subject remains recognizable;
|
|
- the dominant practical source remains dominant;
|
|
- blacks, highlights, and restrained tint remain coherent;
|
|
- movement is deliberate rather than frantic;
|
|
- signal treatment is embedded but does not obscure the subject;
|
|
- no accidental logos, text mutations, anatomy defects, color explosions, or digital
|
|
corruption appear in sampled moments.
|
|
|
|
A sample validates sampled moments only. Record the sampling rule, timestamps, and
|
|
limitations in a concise inspection note, using the `ffmpeg` skill's video inspection
|
|
report when that workflow is active.
|