How AI generation in YouViCo is designed: your project is the prompt

TL;DR: YouViCo’s AI generation is not a separate playground. You start it from the same New file menu as an upload, attach files that are already in the project as references, and the result comes back as a file with comments, drawings, status tags, and a record of how it was made. Before you click, the modal shows the most credits the run can cost; if it fails, they come back. This post walks through those decisions and the reasoning behind them. The guides in the Prompts section, starting with our Louie’s Y2K drama photo, are built on them.

Generation starts where the work already is

Most generation tools are a blank text box on a separate site. You describe your project from scratch, download the result, and upload it somewhere your team can see it. Every round trip loses context.

In YouViCo, generation is the third option in New file, next to YouTube link and Upload. You are already in the project, in the folder where the clip should live, with the brief, the previous cut, and the client’s comments a click away. The prompt field asks one thing: describe what to make. Everything else is a choice from a short list.

The result is a file. Not a preview in a chat window, a file in the project tree, with the same status as any upload: in progress while it generates, then ready for review. Timestamped comments, drawings, and approval tags work on it the moment it is playable.

The AI reads your files before it writes a pixel

The field that matters most is References. You attach up to five files from the project: a storyboard PDF, a reference cut, product stills. Before the generation model runs, YouViCo reads those files and folds what it finds into the request.

This is why the guides tell you to keep prompts short and references rich. A page of storyboard and thirty seconds of graded footage carry more about your look than any paragraph you could type, and they are already in the project. The model gets your palette, your light, and your product’s actual shape, not an approximation of them.

Reading references has a cost, and it is not hidden: adding a reference moves the credit estimate up. Running the same prompt again with the same references is cheaper, because what was already read is not read twice. We wanted iteration, the part of generation that actually produces good work, to get less expensive the more you do it.

One knob for quality

Early versions had a separate quality setting. It duplicated what the model choice already means, so it went away. You pick an output format, video or image, and the modal offers the models available for that format. Each entry says what it is and what it costs. Quality is the model; there is nothing else to tune.

A ceiling, not an estimate

Before you create the file, the modal shows the maximum number of credits the run will reserve. That number is a commitment. If the run costs less, you are charged less. If it costs more than we predicted, the difference is on us. If it fails, the reservation comes back in full, and the failed file stays in the tree with the reason, so you can decide whether to try again.

The modal also shows your two balances separately, plan credits with their expiry and purchased credits, and translates what is left into something concrete: roughly how many seconds of video or how many images the balance covers with the model you selected. We would rather you know what a run means before you start it than find out on an invoice.

Honest waiting

Generation takes minutes, and the server does not know how far along a model is. Rather than draw a progress bar that stops at ninety percent, the waiting screen shows what it can know: the format, the model, the elapsed time, your prompt, and your references, with a cancel button in plain sight. Cancelling charges only the work already done.

When the model finishes, you see the result before the high-quality encode is complete. A quiet bar says a better version is on the way, and the player stays where it is when the swap happens. You can start commenting on the rough version; the comments carry over.

Every result remembers how it was made

Open the file’s info panel and the generation block is there: the model, the output type, the prompt, and the references that were used. When one result is the one the client loved, the recipe is attached to it. The next episode of the series starts from the same prompt and the same references, changing only what should change.

What this means for the guides

Each guide in the Prompts section follows this shape because the product does: put the references in the project, open New file ▸ AI, choose the format and model, paste a short prompt that points at the references by name, read the ceiling, create, and review the result with your team. Start with Louie’s Y2K drama photo guide, then browse the Prompts section for the rest.

If you use generation in a way we have not written up yet, tell us at contact@youvico.com. The best guides come from real projects.