This skill set writes text prompts for someone else's video model

The dependency chain has three links and you own two of them

Read the output list carefully and the first item defines the product. What comes back is a complete prompt of fifteen to twenty-five lines, paste-ready, with a timestamp for every second. Everything else in that list is a section of that prompt: hook engineering for the first two seconds, camera movement direction, lighting and mood presets, sound design layers, platform optimisation, and notes on where to place material references.

No video is generated. The README says so in the setup instructions rather than in the feature list: you paste the prompt into another product and get the result there. The chain is Claude, then these skills, then a video model on a hosted platform.

That third link is a commercial dependency the repository cannot set up for you. You need an account there, and the material reference system, which tells you exactly where to use image, video and audio slots, uses that platform's own reference syntax rather than anything portable.

So the honest description of this project is a prompt-writer specialised to one engine. That is a real thing with real value, because prompt phrasing is model-specific and nobody publishes the good phrasing for a competitor's model. It is also narrower than the repository name suggests, and worth naming before anyone installs it expecting a generator.

One detail about provenance: the author says this is the skill set he uses for the video on an account with a large following, copied out of his own agent configuration directory.

The core contribution is a vocabulary translation table

The clearest argument in the README is a contrast between two phrasings of the same request.

One is a request that the engine reads as a mood. The other specifies a camera move with a distance in feet, an aperture, and a focus pull described as going from foreground to background. The claim is that the second is understood better, and that the engine is a precision instrument responding to camera language, lighting terminology and timing breakdowns rather than to adjectives.

If that is true, the value of these skills is the table behind the claim, and the repository describes its contents in exactly those terms. A camera movement encyclopedia of more than fifteen techniques, each with phrasing written for this engine, including dolly, orbit, whip pan, snap zoom and rack focus. Lighting and mood presets given as full setups with key, fill, rim and practical sources and a colour temperature, in presets such as three-point, neon contrast, silhouette, golden hour and practical. And a four-layer sound design stack of ambient, impact, music and voice.

The layering of specificity is consistent. Each skill claims five or more worked example prompts at the same fifteen-to-twenty-five lines, which is the length the engine is supposed to want.

Whether the phrasing is tuned or merely confident is the question a reader cannot settle from the repository, since there is no evaluation here and no comparison against an untuned prompt. What is verifiable is the shape of the claim and where it points.

Ten skills, three groupings, and numbering that runs across them

The catalogue is organised by what you are trying to achieve rather than by technique, and the numbering runs one to ten across the whole set.

Grouped for views are three: a viral hook skill for scroll-stopping openers using pattern interrupts, curiosity loops and dopamine triggers; a faceless channel skill for producing video with no face and no camera content for the three short-video networks; and a podcast visual skill that turns an audio clip into video, described in the table as killing the audiogram.

Grouped for business are three as well: a product launch skill whose stated target is video that looks like a keynote; a course promo skill for trailers selling coaching programmes; and a testimonial story skill for social proof.

Grouped for brand are four: a founder authority skill, a luxury aesthetic skill pitched at high-end visual quality on no budget, a transformation reveal, and a digital persona skill for producing avatar content at volume.

Because the numbers run across the groups rather than restarting, the table order is not the skill order, and the directory names inside the skills folder are what you actually select on.

The framing is worth noticing as well. Every entry is a content format with an audience and a platform attached, not a prompt technique. The result is closer to a production department than a toolbox.

The first two seconds carry the entire argument

The README's explanation of why short prompts fail leads with retention, and the numbers it uses are worth quoting precisely because they are the author's framing rather than a measurement in this repository.

It says the first two seconds determine whether half the viewers bounce or ninety percent stay. Every prompt opens with a hook engineered for retention, and each skill carries more than ten attention-capture patterns specific to its style.

Sound is treated as part of the hook rather than as decoration. The claim is that sound is half of it, and that every audio element, named as a bass hit, a whoosh, an ASMR texture and silence, is timestamped to a visual beat. That is a specific and unusual instruction, because silence placed deliberately in a timeline is a different craft from adding music.

The timeline architecture underneath it is described as a beat-by-beat breakdown for every second of a clip between four and fifteen seconds. That length range is the real constraint on everything else, and it is worth checking against what you plan to make. A thirty-second explainer is outside what these skills describe.

The opening patterns are named rather than shown in the README, so what you get on install is the taxonomy rather than the techniques. The examples in the usage section show the shape of a result, not its content.

Install means writing into your own dot-directory

There are three installation routes and none of them is a plugin marketplace entry.

The first clones the repository into the agent configuration directory:

cd ~/.claude/skills
git clone https://github.com/rediumvex/ai-video-generator-claude.git ai-video-generator

The second is the same clone one level deeper, for people keeping their skills inside a shared collection directory rather than at the top level.

The third is an installer script inside the repository:

cd ai-video-generator
python install.py          # Pick which skills to install
python install.py --all    # Install everything
python install.py --list   # Preview available skills

After any of them you restart the coding assistant, and each skill inside the skills folder becomes available.

Two things about that arrangement are worth noticing. The installation target is a hidden directory in your home folder, so uninstalling is a directory deletion rather than a package removal, and there is no manifest recording what was installed by which route. And the only executable code in the repository is that installer, despite the project being described by its primary language as Python; everything else is prompt material.

The interactive installer's preview flag is the one to use first, since it lists what is on offer before you commit a directory to it.

Platform tuning is a separate concern from the look

One of the outputs is platform optimisation, specified for four destinations with three specific concerns: aspect ratio, safe zones and loop potential.

That is a different axis from everything else in the skill set. The camera, lighting and sound material is about what the footage looks like; this is about where it will be posted and how the frame has to be composed so nothing important sits under a platform's own interface. Safe zones are the reason a subject cannot simply be centred.

Loop potential deserves a note of its own, because it is the only output in the list that describes a property of the viewer's second viewing rather than of the video itself. A clip designed to loop has an end that matches its beginning, which changes what you can put in the final two seconds and therefore what the hook has to set up.

The usage examples show the skills composing rather than competing. A request for a viral opener for a finance channel with no face on camera pulls in two skills together, one for the hook and one for the format, which is the intended behaviour for a set organised by goal.

The other two examples show the same composition: a product launch request producing a prompt with camera, lighting, sound and timing, and an audio clip becoming a short-form visual story. In each case the README shows what was asked and which skill answered, and then stops.

What the repository contains, and what the author claims for it

The top level of the repository is three files and one directory: the licence, the readme, the installer, and the skills folder. There is no test suite, no continuous-integration configuration, no documentation directory and no release history.

That is a normal shape for a prompt collection. It is worth stating anyway, because the README's central claim is empirical rather than literary, and there is nothing in the repository that would let you test it.

The author's framing is unusually direct about what this is for. It says he builds products that make money rather than demos, and it attributes the pipeline to his own output on an account with a reported following in the hundreds of thousands. Two sibling repositories in the same account cover blog writing and social captions, so the skill-set format is a habit rather than a one-off.

There is also a paid community linked from the readme, offering a workflow automation library, skills, outreach systems and live sessions. That is the business this repository is a demonstration of, and a reader deciding whether to install it should know that before reading the feature list.

The licence is MIT and the copyright line names the author directly. The last push on the repository was on 2026-08-11.

Against asking the video model directly

The alternative is to open the video model, describe what you want, and iterate on the result. That loop costs you time and produces the same class of footage, and its advantage is that you learn the model's vocabulary from the model rather than from a table someone else wrote.

That is the real choice here, and it comes down to how many videos you make. For a handful, the iteration loop teaches you more per hour than reading a prompt framework teaches you, because you are learning the actual response of the actual model on your actual footage.

For volume, the table wins. The camera phrasings, the lighting setups with named sources and colour temperature, the beat-by-beat timeline, the four-layer sound design and the per-platform safe zones are all things you would otherwise rediscover one failure at a time.

The second alternative is a general-purpose prompt-writing skill rather than a video-specific one. That loses the engine-specific phrasing, which is the entire premise, and keeps only the parts you could have written yourself.

There is a third option the repository does not discuss: writing your own template with your own structure and your own examples. For a format you shoot weekly, a saved document beats ten installed skills, because it accumulates the phrasings that worked on your content rather than on someone else's.

Editorial conclusion

These skills suit someone who already pays for the video model they target and has watched the same two-line prompt return generic footage, because the whole contribution is a longer, more specific prompt that the model responds to. It does not suit someone who wants to generate video locally, or someone who plans to switch models, since the camera phrasing and material reference syntax are written for one engine. Verify two things before you install: that the model and platform named are the ones you use, and that a fifteen-to-twenty-five-line prompt is something you will actually paste rather than a starting point you rewrite. The licence is MIT, the last push on the repository was on 2026-08-11, and there are no published releases.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. rediumvex/ai-video-generator-claude on GitHub