Numbers you can compose to
Every measurement here carries its source and the date it was read. Where nobody publishes a number, this section says so instead of inventing one.
Model capability matrix
Which video models generate sound, expose a negative prompt, accept a start or end frame, ship open weights and offer an API. 68 models, each cell carrying the vendor page it came from, and "not established" kept distinct from "no".
Why AI video looks wrong
40 generated video defects mapped to the cause and the prompt change that fixes each one. Extra limbs, morphing, drifting characters, text nobody asked for, invented audio, and the ones specific to commercial and animated work.
How long should a clip be
What a failed generation costs at each length, how to estimate the seconds a scene needs from its word count, what forces a cut and what does not, and the longest single generation each model offers.
Animation vocabulary
44 animation and motion terms, each with what it changes in a generated clip and a clause to paste. Includes the live action vocabulary that fights an animated style rather than helping it.
Prompt length limits by model
How long a prompt each video model accepts, what its vendor actually recommends, and why those two numbers are so far apart. Covers hosted APIs and the text encoder token limits on models you run yourself.
Platform safe zones
Vertical video safe areas for 9 platforms, 6 of them with numbers from official sources. Includes an overlay you can switch between platforms, the permanent-versus-conditional distinction most guides skip, and the disagreements between sources recorded rather than resolved.
Prompt craft
The vocabulary that actually moves a generation, organised by the decision it controls, for video, image and audio. Plus a cinematic-language reference of lenses, stocks, lighting and mood recipes, and what Anthropic and Google say in their own prompting guides.