Skills that do the job, not describe it
A skill is a folder an agent reads: the procedure in Markdown, the facts in JSON, the deterministic part in a script. These are built from this site's records, so the capability a skill recommends on is the one the vendor documented, with the date it was read.
Download, unzip into your agent's skills folder, restart. No account, no key, no network call at runtime.
Available now
Video model match
Which video model can actually do what you are asking for — and what to do when none of them can.
Why these exist
The hard part of generative video is not the prompt. It is knowing which of sixty-eight video models can accept the thing you already have, and what to do when none of them can. That knowledge sits in vendor documentation that changes monthly, and the answer to “can I use my character’s photo?” depends on a distinction most documentation does not make.
A skill closes that gap by carrying the data and the procedure together. It is the difference between a guide you have to read and a colleague who has read it.
Proposed next
Not built. Each one is matched to a role that already makes these decisions on a set, and to records on this site it could execute against — because a skill that cannot execute is a document with ambitions.
Shot list splitter
The gap. Someone writes a paragraph describing thirty seconds of action and asks one model for it. Models drop the second beat and cap out at a few seconds, so the result is the first idea, badly. Nobody tells them the answer is a shot list.
What it would do. Breaks a description into shots of one action each, sized to the chosen model's documented maximum duration, and emits a numbered list with the continuity notes that keep identity across the cuts.
Executes against: maxDurationSeconds for 42 models where it is established, the video prompt aspect order, and the shot structures in the cinematic guide.
Lighting plan from a mood
The gap. "Make it look cinematic" is not a lighting instruction, and a model reads it as brightness and contrast. A gaffer hears a mood and knows the key, the ratio, the practicals and the colour temperature that produce it.
What it would do. Turns a mood into a specific setup — key position, ratio, practicals, temperature, grade — then writes it as prompt clauses in the right order, with the lens and stock that conventionally belong to it.
Executes against: 14 mood recipes, 10 named lighting styles, 18 lenses and 7 grades, each labelled as craft convention rather than measurement.
Platform framer
The gap. A shot composed centrally loses its subject behind TikTok's caption block and action rail. People discover this after publishing, and the safe-zone numbers most of the web repeats are wrong.
What it would do. Takes the target platform and returns the usable box in exact pixels, the aspect ratio to generate at, and where to place the subject — then checks a described composition against it and says what will be covered.
Executes against: The verified safe-zone dataset: 7 platforms with per-zone confidence levels, permanent-versus-conditional overlays, and recorded disagreements between sources.
Failure diagnoser
The gap. A result comes back wrong and the person cannot tell whether it was their prompt or the model's limit. Research on text-to-image users found that when they cannot tell, they assume it was them, and they stop.
What it would do. Takes a description of what went wrong and returns the likely cause, whether it is fixable by prompting, and what to change — separating "you asked for two actions" from "no current model does permanence".
Executes against: The defect-cause guide, the concept pages on temporal consistency, physics and identity, and each model's documented control surface.
Audio brief writer
The gap. Most video models return no sound, and the ones that do will not take direction on it. People either accept silence or paste a music prompt and hope. A sound designer builds ambience, then spot effects, then music, in that order.
What it would do. Reads a shot description and writes three separate prompts — ambience, spot effects, music — with the documented tags for the model chosen, and flags which tags are vendor-documented and which are community convention.
Executes against: audioNative for 67 models where it is established, 23 documented ElevenLabs tags, 17 Suno structure tags recorded as convention, and the audio prompt aspects.
Prompt craft · Model directory · Capability matrix · Cheatsheets