This skill writes video prompts by deciding who filmed the shot
Hermes skill for realistic AI video prompts for Seedance and text-to-video models.
At a glance
- What is it?
- An agent skill that generates AI video prompts by first fixing the footage's device, era and photographer, then writing the resulting flaws and a beat timeline into a four-section prompt. The method is the product, and the documentation is Chinese only.
- Who is it for?
- This skill suits anyone generating video whose results keep looking like advertisements when they wanted footage that reads as unstaged, and its four-section structure with an explicit identity-consistency clause and camera exclusions is worth adopting whichever agent you run. Stay with aesthetic prompting when a polished look is actually what you want, since that is what those adjectives reliably produce.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 45 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
Deciding who shot the footage before describing it
This is an agent skill for writing prompts for AI video models, and its central idea is a reframing rather than a technique. Instead of describing how a shot should look, it first establishes where the footage supposedly came from.
The README puts the contrast directly. A conventional prompt asks for a cinematic morning street with beautiful light. This skill instead settles that the clip is a commuter's vertical phone video taken at a Manhattan corner in the 2020s, and only then writes the prompt.
That single decision cascades into everything else. Once the footage has an owner, a device and a decade, the device's limitations follow: the sensor's oversharpening, the vertical framing, the exposure stumbling as the walker passes from shadow into sun, the bob of someone filming while moving. Those artefacts are what the eye reads as real, and no amount of asking for realism produces them, because they are flaws rather than qualities.
The audience is anyone generating video who has noticed their output looks like an advertisement when they wanted it to look like something that happened.
Four sections, and the one that stops the model wandering
The generated prompts follow a fixed structure, and each part does a job.
A character section describes the person in detail, down to fabric, hair behaviour in wind and visible tiredness, and closes with an explicit instruction to maintain the same identity, clothing, hairstyle and appearance throughout the video. That last clause is the practical one. Identity drift across frames is a characteristic failure of generated video, and stating consistency as a requirement rather than hoping for it is the cheapest available defence.
A location section lists concrete objects rather than atmosphere: wet asphalt, taxis, a steaming manhole cover, a breakfast cart, the stairs of a subway entrance. Nameable objects give a model something to place; a beautiful morning gives it nothing.
A camera section states the format and the flaws, and importantly states the negatives: no horizontal framing, no tripod stabilisation, no professional lighting. Those exclusions are doing the real work, since a video model's default is exactly the polished look being avoided.
A timeline breaks the clip into roughly two-second beats, saying what happens in each. That converts a prompt from a description into a sequence, which is what a video actually is.
Device packages, era by era
The second worked example shows the method's range. Asked for a Korean neighbourhood in the afternoon with a home-video feel, the skill switches to what the README calls a consumer video camera package from the 2000s.
That is the idea generalised: a decade and a device imply a whole bundle of characteristics, and swapping the bundle changes everything downstream without the user specifying any of it. A period camcorder brings its own resolution, colour response, focus behaviour and handling, and asking for that package is far more economical than listing the artefacts individually.
The repository backs this with reference documents rather than leaving the packages implicit. Alongside the main skill file sit three references covering camera aesthetics, an atmosphere dictionary, and an anti-artificiality checklist.
The last of those is the most interesting inclusion. A checklist for removing the signals that mark video as generated is a different artifact from a prompt guide: it is a review step, applied after drafting, and it matches one of the documented usage patterns where you hand the skill an existing prompt and ask for it to be made more real.
Three ways to use it, including not installing it
Installation into one supported agent is a single clone into that agent's skills directory.
git clone https://github.com/zhouwei713/seedance-prompt.git \
~/.hermes/skills/creative/realistic-video-promptingAfter a restart or a new session the skill is available by name. For other environments, the README says to place the skill file directly into whichever agent supports skills or long prompts, naming several.
The third option is the honest one and it is listed as a first-class choice: read the files as a prompting methodology and use none of the tooling. The README names the skill file and the three references as the things to read.
Offering that path tells you where the value sits. This is a written method, not software, and the packaging is a convenience. A reader who takes the structure, the four sections, the device-package idea and the exclusion lists, has most of what the repository offers regardless of which agent they run.
Usage is conversational, with documented examples covering a scene with a duration, a scene with an era and device feel, a mood request, and a request to revise an existing prompt toward realism.
What it cannot do
The first limitation is structural: a prompt cannot exceed the model behind it. Asking for exposure to stumble as a walker crosses from shadow into sunlight only helps if the video model can render that. On a model that cannot, the instruction is wasted tokens, and the skill has no way to know the difference.
Related, prompts of this length and specificity may exceed what some models accept or reliably follow. A four-section prompt with a beat-by-beat timeline is a lot of instruction, and models vary in how much of it survives. The README does not discuss per-model limits, and the badge listing several target models suggests breadth rather than tuning for any one.
The documentation is written in Chinese throughout. The method transfers, and the examples are the most useful part, so a non-Chinese reader will be working through translation for the parts that matter most.
The repository is small with no releases and the last push on 2026-08-04, which for a written method is less concerning than it would be for code, since a prompting technique does not rot the way a dependency does. It does mean pinning nothing and expecting nothing.
Finally, realism is a style like any other. A prompt that successfully produces convincing amateur footage produces exactly that, which is worth being deliberate about in any context where a viewer might take generated video for a record of something.
Aesthetic prompting is the alternative, and specificity is the difference
The alternative is what nearly everyone does: describe the desired look with adjectives, asking for cinematic quality, high resolution and attractive lighting.
The difference is what the model is being told. Aesthetic adjectives push toward the polished centre of the training distribution, which is why that approach reliably produces something resembling an advertisement. It is fast, needs no method, and is fine when an advertisement is what you want.
Provenance specification pushes the other way, toward a specific corner: this device, this decade, this person, these flaws. It takes far longer to write, which is precisely why packaging it as a skill makes sense, and it is the only approach that reaches footage which reads as unstaged.
Take the adjectives when you want something that looks good. Take this method when you want something that looks like it happened, and read the anti-artificiality checklist even if you never install the skill, since a review pass for the signals that give generated video away is useful against any prompting approach.
Editorial conclusion
This skill suits anyone generating video whose results keep looking like advertisements when they wanted footage that reads as unstaged, and its four-section structure with an explicit identity-consistency clause and camera exclusions is worth adopting whichever agent you run. Stay with aesthetic prompting when a polished look is actually what you want, since that is what those adjectives reliably produce. Take the third installation option the README offers and simply read the skill file and its three references, because the value here is a written method rather than software, and note that a prompt asking for device flaws only helps if the video model behind it can render them.
Frequently asked questions
What makes this different from ordinary video prompting?
It establishes the footage's origin first, deciding who filmed it, on what device and in which decade, then writes the device's limitations, natural mistakes, environmental sound and a timeline into the prompt, rather than describing a desired look with adjectives.
What structure do the generated prompts follow?
Four sections: a detailed character description closing with an instruction to keep identity, clothing and appearance consistent throughout; a location listing concrete objects; a camera section stating format, flaws and explicit exclusions; and a timeline broken into roughly two-second beats.
Do I have to install this into an agent?
No. The README lists three options: cloning into one supported agent's skills directory, placing the skill file into any agent supporting skills or long prompts, or simply reading the skill file and its three reference documents as a prompting methodology.
What is in the references directory?
Three documents covering camera aesthetics, an atmosphere dictionary, and an anti-artificiality checklist. The checklist supports a documented usage pattern where an existing prompt is handed over and revised to remove the signals that mark video as generated.
Community notes