Making a Short AI Film in Topview Canvas With the Astra Agent

OpenAI's GPT 6 Astra model has drawn attention for being expensive in tokens and strong in many areas, including research, writing and work inside third-party tools such as a desktop compositing app and the browser-based design platform Topview AI. The compute cost makes sense once the output quality is visible. The more interesting shift is an LLM acting as an autonomous agent inside a video editor, which removes the need to bounce between five tabs to produce one usable clip.

What follows covers what is different about Astra, how to select it as the agent in Topview's Canvas tool, and a full short-film example from prompt to finished video.

What changed in Astra

OpenAI did more than widen the context window. The model understands assigned tasks differently, in three ways that matter for video work.

• Native visual understanding. Older models translated text into images and hoped for the best. Astra handles everything in a unified way, so reading a character description gives it a sense of that character's spatial dimensions.
• Proactive planning. Instead of answering immediately, it behaves like a director: it studies the script, works out the camera angles it needs and builds a logical shot sequence before writing any video prompts.
• Memory across a long project. In a three-minute video, the last scene has to remember the first. Astra tracks details such as a character leaving a jacket behind in one shot and not wearing it in the next.

Setting it up in Topview

Topview offers advanced video generation tools and supports current models, including Seedance 2.5 for video and GPT 6 Astra for agents. Many platforms make users assemble the pieces by hand, whereas Topview built a whole ecosystem around the generation process.

Start by creating an account at topview.ai. The demo here uses the Canvas tool to make a mini AI film, with the agent's model set to GPT 6 Astra in the dropdown. Choosing the right model matters, since that is what supplies the planning ability described above.

For the story, the Short Film Scriptwriter skill is combined with a prompt. The one used: a story about an imaginative eight-year-old named Leo and his scruffy, patient rescue mutt Barnaby, set on a crisp autumn afternoon in the woods behind their house, where the two play at being brave knights hunting for treasure. It stresses the quiet, unspoken bond between boy and dog, as Barnaby fetches magic sticks and guards a cardboard-box castle, and ends by showing how a day of make-believe captures the comforting magic of a child's friendship with a pet.

Canvas then opens a new workspace and the agent begins planning how story and media assets fit together. It asks a series of questions, which should be answered in the floating chat panel on the right. This step matters. The agent wants the intended pacing and emotional tone, and good answers here save considerable editing later.

As the answers accumulate and the story sharpens, scenes appear as nodes in the center of the screen. Watching them link up in real time offers a clear picture of how the narrative is being structured.

What improved after switching to Astra

After months of using Canvas, the notable differences from switching the agent to Astra are these:

1. Better scenes and dialogue. Stories improved in quality. The agent asks sharper questions, so there is less to tweak, and dialogue sounds like something a person would say instead of the stiff, robotic lines typical of earlier models.
2. Faster generation. Quick story drafts keep creative momentum steady and allow more videos. When a script draft does not take ten minutes, trying wildly different concepts becomes practical.
3. More control. Astra asks whether the project should be a micro film, a short emotional story or a full feature film. Earlier versions did not ask this. It works like an assistant director anticipating what the project needs.

Next come storyline confirmation, script writing and finally scene generation. A scene produced by Astra looks like this:

Scene: late afternoon EXT cardboard castle in Autumn Woods Behind the House

Cast: Leo — Woodland Knight ; Barnaby — Patient Rescue Mutt


Use Leo — Woodland Knight
 for eight-year-old Leo's face, tousled brown hair, moss-green sweater, brown trousers and boots, and russet play cape. Use Barnaby — Patient Rescue Mutt
 for his rescue dog's medium-sized build, scruffy tan-and-charcoal coat, asymmetrical ears, and collar. The character sheets supply appearance, not panel layouts. Autumn Woods Behind the House
 supplies the leaf-covered clearing, oak, homeward path, geography, and autumn daylight, now gently later in the same afternoon. Leo and Barnaby's Cardboard Castle
 supplies the brown box den, broad arched entrance, doorway flap, tape repairs, right-hand square window, crayon shields, and uneven battlements. The castle stands on the level patch beside the oak as in the earlier memory; ignore its reference image's plain background. Leo brings the same ordinary brown bark-covered crooked stick, forked at one end, back from the quest.

The nodes on screen then show each scene linked in order, and each scene can be generated with Seedance 2.5. After the first scenes are done, the agent can generate the rest and stitch everything into the final video, leaving the platform to do the heavy lifting.

The finished film took under 30 minutes with Astra and Seedance 2.5. Facial expressions fit the dialogue well, scene transitions work, every scene looks consistent with the story, and the short piece is well executed. For a human artist without such models, the same result would take days: rendering each frame, checking lighting and adjusting character rigs by hand to get five seconds of usable footage.

Getting more from the questions

The agent's opening questions deserve deliberate answers, because they steer everything downstream. A few habits follow from the way Canvas behaves:

• State the format early. Choosing between a micro film, a short emotional story and a feature changes how many scenes the agent plans.
• Describe the feeling as well as the plot. Tone and pacing answers shape dialogue and shot rhythm, and vague replies lead to generic scenes.
• Keep the character descriptions specific. Consistency across scenes depends on how clearly each character is defined at the start, which is what the scene text above relies on.
• Review the node layout before generating. Reordering or rewording a scene is cheaper at the planning stage than after rendering.

Why this matters

A few years ago, an animated short of two minutes, with the same two characters, a woodland setting in autumn and lighting that matched from shot to shot, needed a small team and several weeks: a storyboard artist, a character designer, an animator and an editor. Most people with a story idea never made it, because the gap between idea and skill was too wide to cross over a weekend. The slow, manual stages have mostly vanished, and the implications are not yet fully absorbed.

• For filmmakers, the biggest gain is iteration. A version of a scene can be tried, judged wrong and replaced with another in the time it once took to book a session. Failing with an idea no longer costs much.
• For content creators and marketers, the pipeline shortens sharply. There are no handoffs between scriptwriter, asset maker and editor. Whether that is good depends on where someone sits, and it is understandable that it makes some people nervous.

One of the most enjoyable and simple uses is the one-shot short film: a parent making a film starring their own child and their own dog. Personal storytelling used to be limited to what could be photographed or written down. Now a bedtime story invented by a child can become a film with a character resembling the real pet.

None of this argues for handing every creative workflow to AI. Human taste still matters. Sensing what makes a scene work, where to cut and why one shot should follow another still separates a clip from something people want to watch. Speed arrived with the tools, but taste was not automated.

Other places the same agent runs

Canvas is not the only Topview tool with Astra integrated. The same agent powers several others, each suited to a different kind of project.

• Drama Studio targets longer formats: short dramas, micro-dramas, vertical dramas and episode-style story videos drawn from a novel, a script or just a rough plot. Before anything is generated, the agent lays out the cast, scenes, props, prompts, continuity and dialogue. It fits when the goal is producing several connected episodes efficiently and not perfecting one shot.
• Film Studio goes the other way. It is the cinematic suite, right when a project depends on exact framing, performances and the connection between shots. Both studios run on Seedance 2.5, so output quality is the same. The difference is the amount of control the interface offers.
• The Marketing Video Agent is where Astra's research strength pays off. Paste a product URL, and the agent researches the product, plans scenes, picks models, generates assets and keeps everything editable on Canvas. Because a model that researches well does something different from one that writes ad copy well, it is useful for brainstorming campaign angles.
• Motion Studio is the newest, pitched as a way to make motion videos with AI in place of a desktop compositing app, with the same agent figuring out how to animate something for a brand or product.

Canvas also contains a 3D Shot Composer for blocking camera angles before generating anything, and Clip Edit, which changes one section of a video while leaving the rest intact. Both are worth exploring once the basics feel comfortable.

Recap of the workflow

In order, the process was: pick Astra as the agent, respond to its questions about pacing and tone, and watch a one-paragraph premise about a boy and a dog grow into a storyline and then a script. After that came scene descriptions specific enough to hold both characters steady across shots, and finally Seedance 2.5 renders that the platform joined by itself. A browser tab was the only software involved, and the total stayed below half an hour. The gap separating a story idea from a finished film used to be enormous, and seeing it shrink this quickly is among the more striking changes in the field.