AI Video Prompt Craft — Kling 3.0 & Seedance 2.0
The prompt grammar and cost routing behind a working AI video operation.
The prompt grammar and cost routing behind a working AI video operation.
What is in it
- skills/ — 14 files
- LICENSE.md — included in the package
- README.md — read this first
Claims, and how to check them
- the archive contains 18 files
check: list the zip; this count was read from it
What it does not do
- not a program — nothing here runs on its own
- generation credits are your cost
- no guarantee of an outcome, and no revenue is promised
You need already
- nothing to read it
- your own generation credits to run anything in it
files: 18formats: md, jpg, py
The prompt craft from a working AI video operation — every rule came out of a failed order that cost real credits.
This is not a vocabulary list. It is the set of things vendor documentation will not tell you.
WHAT'S IN HERE THAT ISN'T IN THE DOCS
A negation is not a ban — it's a summon. Write "no dust" and you get dust. The pack has a one-line test for when negation works and when it backfires.
Some models silently ignore the resolution parameter. You raise the mode, not the resolution field. Not knowing this means forever wondering why you ordered 1080p and got 720p.
A 5.1x price spread on the same 15 seconds. Which shot goes to which model is your cost. Full three-tier price table included — measured, not estimated.
Close-up plus large rotation equals face collapse. Two motion axes per shot, maximum.
Only wide shots drift into a different world — a subject lock does not lock the look.
A half-hidden face reads as horror. Three triggers, and the fix.
Travel direction is set by the starting image, not by wording.
THE 11 SKILLS
prompt-pipeline (the map) · prompt-model-routing (constants SSOT — model choice is cost) · prompt-kling · prompt-seedance · prompt-camera-motion · prompt-subject-motion · prompt-expression-acting · prompt-consistency-lock · prompt-lipsync (plus a working reference-builder script) · prompt-moderation · prompt-troubleshooting (24 symptoms, cause and action)
TESTED BY BUILDING, NOT BY READING
Before release, buyers-from-scratch installed this pack and actually produced finished videos with it, spending real credits. Every defect they hit was fixed before this went on sale. Specification numbers were cross-checked field by field against the live vendor API.
BEING STRAIGHT WITH YOU
These are observations from our account. Models invert behaviour between versions — so instead of claiming a fixed ranking, every skill ships with a re-measurement procedure.
Model names are not baked into the rules. Criteria are written as properties (high-quality mode, "a model that accepts video references"), so the pack survives a service change.
Where current models simply cannot do something — simultaneous lip sync for two people in one frame — it says so, and gives you two workarounds.
No promises about views or revenue.
REQUIREMENTS
Claude Code · access to an i2v/t2v generator (Kling or Seedance family) · your own generation credits. ffmpeg and Python 3.9+ only for the lip sync helper script. No web search needed. No specific subscription required.
LICENCE
Personal and in-house use included. Agency use (delivering to third-party clients) requires a separate licence. Reselling the source text is prohibited. Commercial use of the video you make is entirely yours.
Lifetime updates.