How to work with CinAI Skills

CinAI runs image, video and audio models on a drag-and-drop canvas. You wire up inputs, a model runs, and the result lands back on the canvas where you can feed it into the next step. Credits are pay-as-you-go — you spend them when a model actually runs.

This site is the public library for that studio. It lists the skills and the workflows so you can see what the system produces before you sign in anywhere.

A skill is not a prompt. It is a saved graph: the models, the inputs they expect, the prompts already written, and the order they run in. You open it, put your own material in, and run it.

Every skill page carries two pieces of text written by the team that built it. One says what the skill produces. The other says when to reach for it — which is usually the part you need, because picking the right workflow matters more than tuning any single prompt.

A workflow is a graph of nodes joined by connections. Some nodes hold what you put in — a reference image, a video, a piece of text, an audio clip. The rest are runs: a node calls one model with a prompt and produces an image or a video.

The connections are the important part. A generated frame can become the reference image for the next run, so character and scene stay consistent across a sequence instead of drifting shot to shot. That is why these are graphs and not a list of prompts.

The library is grouped by the job the work does, because that is how buyers describe what they want. Nobody asks for an image-to-video model. They ask for a product video, an explainer, a character that stays the same across ten shots.

So start from the deliverable. Pick one category, learn the two or three workflows inside it well enough to run them on someone else's material, and you have something specific to offer. A narrow niche you can demo beats a broad one you can only describe.

The studio runs several families of model: image generation, video generation, video editing, audio, text, storyboarding and 3D. A workflow already names the model on each node, so most of the time you do not choose anything.

When you do swap one, swap it for a model in the same family and change one node at a time. A different video model with the same prompt can shift the whole look of a sequence, which is easy to miss until you compare two runs side by side.

The program terms, tracking and payouts live on the program page — that page is the source of truth for anything about money, so check it there rather than taking a number from a docs page.

What this page can help with is the part that decides whether promoting works at all: having something specific to show. Every skill and workflow page here is a public URL with real output on it, so you can link straight to the thing that answers a question instead of to a home page.

Double-click empty canvas to create a node. You can also drag an image, video or audio file straight onto the canvas from your desktop, and it becomes a node of the matching type.

The source guide keeps its earlier script implementation documented alongside the current one. It is marked as superseded there, so treat anything in it as history rather than instruction.

With text, image, video and audio nodes on the canvas, you draw connections between them to build the workflow you need. Once it is built you can run the whole thing at once, or hand it to someone else to run.

Four panels sit around the canvas: the project menu, the canvas sidebar, the account area, and the minimap.

Top left of the canvas. It holds the project-level actions — reaching all your projects, and the settings that apply to the project you have open rather than to a single node.

The sidebar on the left of the canvas is the library for one project. It is where the assets you saved from nodes are listed, and where saved workflows can be opened and sent onto the canvas.

It is worth using early. The alternative — keeping reference material outside the tool and re-uploading it per shot — is the thing that makes long pieces slow.

Instead of drawing a connection for every reference, you can refer to material by name from inside a prompt and the canvas resolves it to the node that holds it.

On a long piece this is the difference between a readable canvas and a thicket of crossing lines.

These are reached from a slash menu on the canvas. Each one takes a reference and produces a specific named result — a grid of angles, a turnaround, a relit frame — rather than a general image.

The point of them is repetition. Building a nine-angle grid by wiring nine generators is the same work every time, so it became one action.

Produces a storyboard from what you give it, so the shot order exists before any single shot is finished.

The same idea as the plain storyboard, with staging attached — who moves where, and where the camera goes.

Adjusts a portrait — the framing, the look, the presentation — while keeping the person recognisable across the shots you already made.

Generate a 720° panorama from text or a reference image, preview it live, and take screenshots from inside it. The reason to bother: it makes scene consistency easy — every shot of that location comes from the same panorama instead of being generated separately and drifting.

Produces a nine-cell grid from the current image. The grid-splitting tool below is the other half of this: generate the grid, then extract the cell you want.

Use a brush or a box to mark a specific region of the image, then add text inside that region.

The annotation is saved with the image, so regeneration follows the marked position and the text without you writing a long description of where in the frame you meant.

These act on a video node. Between them they cover the whole span from cleaning up one clip to assembling a finished cut, which is the reason the canvas can hold a complete piece rather than just the generation step.

Upscales a single video 2×, 4× or 6×, and raises frame rate to 30, 60 or 90fps.

A slow-down strength parameter controls the slow-motion effect: the higher the value, the slower the motion.

Analyses one video into its shots and presents the result as a table. Opening the table full screen shows the complete breakdown: what is in frame, the shot size, an image prompt, a camera-movement prompt, the start point, the duration, the music rhythm, and where there is speech.

The useful part is the prompts. This turns a reference video into the two prompts you would need to make something like it, instead of leaving you to describe it from scratch.

Basic editing on a single video on the canvas. Currently that means taking a segment out of it.

Separates the picture from the audio. Select the clip on the canvas and choose the split option from the top toolbar.

This is the tool aimed at the repetitive half of editing. It analyses and arranges material across tracks, matches motion, beat and layout to what is in frame, and produces a cut in one action.

Pick the point to continue from, describe what happens next, and optionally add new reference images or video. The continuation is built on the original's characters, setting and visual style rather than starting fresh.

Being able to start from any timestamp is the part that matters: a shot that goes wrong halfway does not mean regenerating the whole clip.

Marks a segment of an existing video and generates a replacement for just that stretch, leaving the rest intact.

A lightweight 3D composition node. You build the space the way you would with blocks — figures, simple shapes, a camera — then take screenshots from the angles you want.

The point is not to render anything in 3D. It is to produce a composition reference that an image or video model reads accurately, which is much harder to get from a text description of where everyone is standing.

All three are reached the same way: select an audio clip on the canvas and pick the operation from the top toolbar.

Precise trimming of an audio clip, to keep the part that is worth keeping.

Speeds audio up or slows it down, so it sits against the rhythm of the picture.

This works across one image or several, which is the point: the character from one frame and the product from another can go into the same new image without you describing either of them in words.

The source's own comparison keeps the prompt fixed and changes only the optics: 125mm at f/11, then 125mm at f/4, then 35mm at f/4. Same subject, three different photographs — which is the argument for the feature. Depth of field and compression are things a text prompt describes badly and a focal length and aperture describe exactly.

The source guide keeps a long table of its own image models with a capability summary for each. That list is deliberately not reproduced here: CinAI publishes its own model catalogue, and two lists answering the same question on one site go out of sync the first time either side updates.

What does carry over is the shape of the differences worth checking before you pick one.

The language models in a canvas tool are not there for chat. They do four jobs: write prompts, read an existing image or video back into a prompt, draft and develop a story or script, and build character definitions.

The multi-modal ones take images, video and audio as instruction alongside text, which is what makes reading a reference back into a prompt possible.

Record a voice sample and it becomes a voice you own in the library. After that, text you enter is generated in that voice, so narration stays consistent across episodes instead of shifting each time.

The shortcut panel opens from the project menu, and the source groups the bindings into creation, zoom, moving the canvas, and everything else.