AI Short Films Aren’t Magic—They’re a Mini Film Crew Moved into Your Computer
Do you think this is an unreleased clip leaked from some sci-fi blockbuster?
In fact, it’s a six-minute AI short film “sculpted” by just the two of us over 12 days—burning through more than 3,000 compute credits.
Outside the industry, there’s a romantic misconception that AI creation is pure magic: utter a spell-like prompt, hit “generate,” and—voilà—a finished film appears. But anyone who’s actually waded in knows the truth: diving in unprepared guarantees disaster.
AI tools are indeed powerful—but they behave more like brilliant yet temperamental actors and cinematographers.
You give your art director a prompt, and they hand you a cyberpunk cityscape. You give your cinematographer a prompt, and they shoot a medieval castle. Without a script supervisor or production coordinator, these fragmented assets simply won’t fit together.
That’s why we didn’t rush to generate visuals right away. Instead, we opened Lark.
Within our two-person mini film crew, Lark became our “digital film set” and collaboration hub:
- Instant messaging kept our creative ideas instantly aligned;
- Documents served as the living repository for scripts and world-building;
- Multi-dimensional tables managed hundreds—or even thousands—of shot assets;
- Calendar and video meetings rigorously enforced our delivery timeline.
Once we’d moved actors, studios, and cameras into the computer, what we were missing most was a clapperboard—one that could command the entire set.
Tools change constantly, but the underlying production mindset remains consistent. Before unleashing those intelligent third-party AI assistants, we first had to erect the scaffolding of our film set.
Don’t rush to generate—let’s first talk about how to build your crew’s “asset library” using Documents and Multi-dimensional Tables.
Don’t Rush to Generate—First Build Your Asset Library with Documents and Multi-dimensional Tables
When starting out with AI short films, the easiest pitfall is “think-as-you-generate.”
The result? Assets scattered across countless chat windows, version chaos so severe that even you can’t recall what your protagonist looked like in Version 1.
To avoid this disaster, we dropped our script directly into a Lark Document.
This wasn’t just typing—it meant breaking down character bios, scene settings, and shot-by-shot scripts into distinct, editable sections. We annotated, debated, and compromised in real time, right inside the document. It’s essentially a miniature version of a traditional writers’ room.
Once the script was locked, the real heavy lifting began: visual asset development.
Characters, locations, props—all were organized inside a Lark Multi-dimensional Table. Each row represented one asset; each column captured its visual traits, emotional tone, even lighting requirements.
Here’s a painful lesson that sent us back to the drawing board multiple times.
Initially, our protagonist’s outfit included unmistakable design cues from a well-known brand. During subsequent generation, our third-party AI assistant triggered an automatic copyright block—and even when images slipped through, they always carried strange distortions.
No choice but to return to the Multi-dimensional Table and manually audit hundreds of asset descriptions.
Only after replacing every instance of “specific brand” with “distressed leather evoking a post-apocalyptic aesthetic” did the generation pipeline finally run smoothly.
An asset library is never a static spreadsheet—it’s a meticulously organized prop room for your film crew.
Now that the prop room is sorted and the script polished, it’s time to bring in smart workflow Agents to “direct the scene” for us—in group chat.
Deconstruct Prompts Second-by-Second: Let Smart Workflow Agents Direct in Group Chat
We invited a third-party smart workflow Agent directly into our Lark group chat.
Let’s single out a heart-wrenching lesson: previously, with scattered AI tools, context collapsed constantly. Switch windows, and the AI forgot what your protagonist looked like—forcing us to redraw cards repeatedly. Hundreds of compute credits burned to ash in minutes.
Now, Lark’s instant messaging serves as our “director’s monitor.”
Within the group chat, the Agent seamlessly inherits all previously established context. We no longer dump long, monolithic text prompts blindly. Instead, we deconstruct each prompt into a three-layer structure—second by second:
- Overview: Sets the scene’s atmosphere and core action;
- Timeline: Breaks down camera movement—push, pull, pan, tilt—frame by frame;
- Constraints: Locks in lighting, aspect ratio, and elements strictly forbidden from appearing.
It feels like a director briefing both actor and cinematographer simultaneously—clear on blocking, lighting, and timing.
“The push-in at second three is too fast—switch to a slow lateral dolly.”
“The lighting is off—refer to the ‘post-apocalyptic dusk’ setting in the Multi-dimensional Table.”
We tag the Agent line by line in the group chat: “@Agent, adjust lighting per Table Row #47.” Instantly, it replies in the same thread with revised instructions and test frames. Our two-person team never switches tabs—the full history of iterations and sparks of inspiration naturally settles into Lark’s chat logs.
Translating abstract imagination into a clapperboard AI can actually read.
Once test frames in chat perfectly match the script, batch generation flows effortlessly.
But staring at hundreds of newly generated video clips piling up on the hard drive brings fresh anxiety: faced with this avalanche of fragments, how can a two-person team efficiently screen them—and hold firm to the editing rhythm?
Generation Is Just the Beginning: Use Calendar and Video Meetings to Enforce Editing Discipline
Hundreds of AI-generated video clips sit on the hard drive—impressive in volume, exhausting in execution.
A two-person team’s biggest enemy has never been technical barriers—it’s aesthetic fatigue and schedule derailment caused by overwhelming fragmentation. When juggling multiple threads, tweaking transitions today and swapping sound effects tomorrow easily traps you in an endless loop of “micro-adjustments.”
Let’s face a harsh truth: AI generates raw material—human editors shape the final film.
Editing by gut feeling is amateur territory—even a micro crew needs professional discipline. So we forced our entire post-production workflow into the Lark Calendar.
Rough cut, fine cut, sound design integration, final review—every critical milestone became a calendar event.
At review time, even if we felt the lighting in one shot wasn’t quite perfect, we paused to assess the sequence as a whole. The Calendar’s strong reminders literally yanked us out of ceaseless “detail-oriented burnout.”
Stuck in a deadlock over whether to keep or cut a particular shot? Then fire up a Lark video meeting.
Don’t laugh—two people still need ceremony around meetings.
Through screen sharing, we scrubbed the timeline to the contested frame and analyzed it frame by frame. One person gestured rhythmically over the shared timeline; the other logged revision notes in real time inside the Document. Built-in meeting recording automatically turned these “debates” into reviewable records—so we wouldn’t wake up the next day wondering why we made that edit.
When progress is bounded by the Calendar and disagreements resolved in meetings, editing rhythm emerges naturally.
AI handles boundless imagination—humans handle grounded execution.
Watching fragments on the timeline gradually lock into place as a coherent story delivers a tangible sense of control—something no “one-click generation” illusion can replicate.
Models will keep updating—but the workflows embedded in Lark won’t.
Models Will Keep Updating—But the Workflows Embedded in Lark Won’t
Twelve days. Two people. Over 3,000 compute credits consumed.
When the final six-minute short rendered and exported, we both posted a long, exhausted “slumped-over” emoji in the Lark group chat.
The fatigue was real—but so was the deep satisfaction of assembling fragments into a cohesive story.
Looking back on this journey, our strongest impression wasn’t how magical AI is—but how rapidly it evolves.
The third-party AI assistant we integrated today may be replaced tomorrow by a smarter workflow Agent. Those prompt-engineering techniques that once drove us mad may become obsolete with the next model update.
Chasing tools alone leaves creators perpetually breathless.
But some things never expire.
Like the shot asset library we built in Multi-dimensional Tables. Like the script structure we rewrote and refined repeatedly in Documents. Like the editing schedule we fought tooth-and-nail to lock down in Calendar.
These collaborative habits—and the finely tuned human-AI coordination—embedded in Lark constitute the only “film crew experience” we truly carry forward.
Tools change—but the underlying production mindset stays the same.
When you treat AI not as a black-box image generator, but as a miniature film crew requiring direction and coordination, you reclaim creative sovereignty.
Accumulated capability will always outpace tool iteration.
That’s why we wrote this retrospective. Building the digital film set has only just begun—and we’re already planning more complex narratives.
If you, too, are intrigued by this kind of “human-AI collaboration”—unafraid of missteps, even enjoying the process of “directing” AI in real-time messaging—
We welcome you to share your work. Our Lark interview calendar is open—anytime.