We were launching Moonshot, an agent that listens to you 24/7, builds deep context about your life, and acts before you ask. The product is futuristic and genuinely useful. It is also, if you phrase it badly, a little terrifying.
That last part is where the creative opportunity was.
The obvious move would have been another polished AI launch film: glowing interfaces, slow product shots, a voiceover about the future of personal intelligence. Instead we built the launch around a much more uncomfortable idea. We said we had built 1984’s Big Brother, and the video crossed 300K+ views.
The interesting part isn’t that Astra helped make it. The interesting part is how we used Astra to get from a product with a complicated story to a creative idea that fit into four words, and then turned that idea into a film. The path looked like this:
Product truth → creative tension → cultural frame → narrative → script → scene direction → generation → critique → diagnosis → iteration → launch
There is a large gap between asking a model to make a video and using one to think through what video is worth making. Almost everything below sits in that gap.
1. The product handed us the creative problem
Moonshot’s promise is short. It listens continuously, it builds context about your life, and it acts before you ask.
Most assistants run on a familiar loop: you have a question, you type it, the model answers. Moonshot changes the timing of that relationship. It isn’t just that it can answer you, it’s that it can hold enough context to notice something before you think to ask.
That creates a hard explanation problem. How do you describe an agent that continuously understands your life without burning the first thirty seconds on memory, context windows, proactive agents, and every other phrase that makes a launch video sound like a documentation page?
We could have made the safe film. Beautiful interface, dark background, AI voiceover, a few product shots, something about reimagining personal intelligence. It would have looked good and been forgotten in a day.
The product had a stronger tension sitting inside it. An agent that listens continuously has to know a lot about you, and that is exactly what makes it useful. It is also exactly what makes people uneasy. Rather than sanding that discomfort down, we made it the opening line.
2. Why the “Big Brother” frame earned its place
“Big Brother” does an enormous amount of communication before the viewer has seen a single feature. You already understand surveillance, someone watching, someone knowing more about you than you realise, and what it feels like to lose privacy. All of that arrives for free, which meant we didn’t have to spend screen time building it.
But a reference only works if the product earns it. Moonshot is not Orwell’s Big Brother, and pretending otherwise would have collapsed the moment the product appeared. The point was never the equivalence, it was the tension. Big Brother knows everything about you. The uncomfortable question is what happens when your personal AI does something that sounds superficially similar.
The difference is what that knowledge gets used for, and that gave us the contrast the whole film runs on:
Surveillance as control vs. context as assistance.
That is far more interesting than trying to convince people an AI assistant is “the future.” It also leaves the viewer holding a question: wait, what exactly did they build? That question was the beginning of the video.
3. We didn’t start by asking Astra to make a video
This is probably the most important part of the workflow, and the easiest one to skip.
The fastest way to get mediocre AI output is to start with the output. “Make a cinematic launch video for Moonshot” contains almost nothing useful. It specifies a format and says nothing about the creative problem, so the model fills the gap with the average of every launch video it has ever seen.
Astra’s first job was understanding, not production. We gave it the product context, the launch objective, the positioning, the audience, and the creative territory we were exploring. Then we used it to interrogate the idea:
What is actually unusual about Moonshot?
Which part of the product creates an emotional reaction?
What is uncomfortable about an agent that listens continuously?
Which part of the product is genuinely differentiated, and which part is table stakes?
What would make someone stop scrolling?
What would make this feel like every other generic AI launch?
Which cultural references could compress this idea into something instantly legible?
Does the “1984” framing actually come out of the product, or are we forcing it onto the product?
What does the viewer need to understand before we reveal anything?
That changes the model’s role. It wasn’t decorating an idea, it was helping stress the idea before we spent anything producing it.
The loop we actually ran
Give context → ask for possibilities → challenge them → choose a direction → add more context → repeat.
Simple to describe, and very different in practice from a one-shot brief. We didn’t want Astra to make a creative decision once and disappear. We wanted it inside the reasoning loop, accumulating context as we went.
Product information shaped the creative direction. The creative direction shaped the narrative. The narrative shaped the scenes. The scenes shaped the visual treatment. Then the visual treatment exposed a problem in the narrative, and we went back.
Most AI workflows are drawn as a straight line: prompt, output, done. Creative work almost never behaves that way. The honest shape is: direction, output, diagnosis, correction, output. The model gets better because the human gets more specific about what is broken.
4. We stress-tested the idea before producing anything
Build an explicit idea test into the workflow. Not a score, not “rate this out of 10,” which produces confident numbers and no information. A short set of uncomfortable questions.
The idea stress test
Recognition. Will someone understand the reference immediately, without setup?
Tension. Does the idea create a feeling stronger than mild curiosity?
Product connection. Does the idea come out of the product, or is it bolted on?
Novelty. Could another AI company make this exact video?
Conversation. Would people hold different opinions about it?
Payoff. Does the product actually resolve the tension the hook creates?
Compression. Can someone explain the whole idea in one sentence?
Moonshot scored well across all of it. Big Brother is instantly recognisable. Continuous listening creates real tension. The premise comes directly from what the product does. And people could disagree about it.
That last one matters more than it sounds. A launch does not need everyone to react positively, it needs people to react. “That’s amazing,” “that’s creepy,” “I want this,” “absolutely not,” and “wait, how does that work” are all far more useful than “cool AI product.”
5. Build the message before the shots
Once the territory was clear, the next available mistake was jumping straight into visuals. Instead we broke the narrative into information layers and decided what the viewer learns, in what order.
The message hierarchy
HOOK. Why should I stop scrolling?
QUESTION. What do I now want to understand?
PRODUCT. What did they actually build?
EXPERIENCE. What does this feel like in practice?
PAYOFF. Why is this useful to me?
CTA. What can I do with it right now?
This small step prevents one of the most common failures in AI-generated video. You can generate six beautiful scenes with no relationship to each other. Every shot looks impressive and the film still doesn’t work, because the viewer never knows why they are being shown each thing.
6. The script was a series of passes, not a generation
We didn’t treat the script as one output either. Each pass looks for a different kind of failure, which is the only way to catch problems that hide behind each other.
Narrative. What is the simplest version of this story?
Compression. What are we explaining that we could simply show?
Hook. Does the opening buy us another few seconds of attention?
Product clarity. Would someone who has never heard of Moonshot understand what it does?
Tone. Does this sound like a real launch, or like an AI company writing its own press release?
Contradiction. Is the narration saying one thing while the visuals say another?
Cut. What can disappear without weakening the story?
The cut pass is underrated. AI makes addition cheap, and cheap addition is not the same as improvement. If a scene doesn’t carry information, emotion, or narrative movement, it is making the film heavier. And no amount of polish rescues a weak idea, so don’t spend passes trying.
7. The scene card became the unit of production
We stopped thinking “make the video” and started asking what each scene has to accomplish.
Scene card
Purpose. Why does this scene exist?
Information. What does the viewer learn here?
Emotion. What should the viewer feel?
Visual. What is actually on screen?
Transition. Why does this scene lead into the next one?
Failure condition. How could this become generic, confusing, or unnecessary?
The failure condition is the field that pays for itself. Models are very good at giving you more of something. Ask for “cinematic” and you get cinematic. Ask for “more energy” and you get more movement. If the real problem is that the scene has no reason to exist, neither instruction helps. Defining what wrong looks like before you generate is what makes the later critique specific.
8. Treat references as evidence, not decoration
Handing a model a few reference images and saying “make it look like this” produces surface imitation. The more useful question is why the reference works.
Any single reference contains a stack of separate decisions: composition, pacing, typography, camera language, lighting, texture, editing rhythm, visual density, how the product is presented, and overall tone. Those need pulling apart before any of them are useful to you.
Reference analysis
Reference. What are we actually looking at?
Mechanism. Why does it work?
Transfer. Which part belongs in our film?
Boundary. What should we deliberately avoid copying?
That converts references into creative inputs instead of a moodboard. It matters because AI can reproduce visual characteristics far faster than it can understand why those characteristics worked in the original.
9. The production loop was never prompt → video
This is where Astra earned the most. The real loop:
Direction → Generation → Review → Diagnosis → Specific correction → Regeneration → Comparison → Keep / kill / modify
Diagnosis is the step that carries the loop. “Make this better” gives the model nothing to work with. A useful critique names the mismatch:
The shot is strong, but the viewer doesn’t understand why it’s here.
The transition adds energy without moving the story forward.
The product appears before the viewer has enough context to care.
The surveillance metaphor lands, but the connection to Moonshot’s value is missing.
The narration is explaining something the visual already communicated.
The scene works alone and breaks the pacing of the sequence.
A model can act on any of those. It can’t do much with disappointment.
10. Astra was also the critic
Separating creative review from production review changed the quality of the feedback. Instead of asking “is this good?”, we asked narrower questions and got answers we could act on.
Creative review. Is the idea still interesting?
Narrative review. Does the sequence make sense?
Product review. Is Moonshot understandable to a stranger?
Attention review. Where could the viewer lose interest?
Taste review. Where does this start to feel generic?
A film can be beautiful and strategically useless. It can explain the product perfectly and still be boring. It can have a great hook and a dead payoff. A single overall score hides all three, and separate reviews expose them.
11. Human approval gates stopped the drift
The model will keep generating forever, which is useful right up until it isn’t. Without gates, production becomes a pile of increasingly polished decisions nobody ever validated.
The gates
Is the idea worth producing?
Does the narrative work without any visual polish?
Does every scene have a job?
Does the visual direction support the narrative?
Is the product still clear?
Does the final cut create a reaction?
When a gate fails, you go backwards. You do not compensate for a broken idea with better transitions. When a video feels wrong, find the layer that broke instead of adding polish on top of it. That is probably the single most reusable rule in this whole process.
12. What Astra was good at, and what it wasn’t
Astra was very good at expanding the creative search space quickly. It helped us explore directions, challenge assumptions, structure messy thinking, reason through narrative, spot weak transitions, pressure-test the hook, critique individual scenes, suggest cuts, hold context across the entire process, and shorten the distance between an idea and a version we could actually judge.
There were things it could not own. It couldn’t tell us whether the concept felt right. It couldn’t take responsibility for the creative bet. It couldn’t know whether a visual was tasteful for this brand simply because it could describe the visual accurately. And it couldn’t tell us that a technically impressive shot should die because it was making the film worse.
That stayed the human job, and it got clearer the more we used the model. AI can produce a great deal of motion. That is different from knowing which motion is worth producing.
13. The 300K question
We had no way to know the number in advance. There is no formula that says Big Brother reference plus AI product plus launch video equals 300K views, and claiming otherwise would be pretending to understand distribution better than we do.
What we could assess was whether the premise had the properties that make distribution possible.
The reference was instantly recognisable, so nobody needed a demo to understand the frame. The product created real tension, because an AI that listens 24/7 sounds extremely useful and slightly invasive at the same time. The product itself was unusual, and another chatbot is very hard to make culturally interesting, while an agent that builds enough context to act before you ask is a different proposition entirely. And the concept was easy to repeat. Nobody had to remember the film to pass it on. They could just say: they built an AI that listens to your life and launched it as 1984’s Big Brother.
So the bet was never “this will get 300K.” The bet was that the idea had enough recognition, tension, novelty, and disagreement in it that people might carry it past the original post. The views validated the bet, they didn’t create it.
14. The launch post didn’t have to explain everything
The copy was deliberately thin. We built 1984’s Big Brother. Then the product proposition. Then the beta.
The post opened a curiosity gap, the video answered it, and the product closed the loop. That’s a useful default for launch content generally. The social post doesn’t always need to carry the whole argument. Sometimes its only job is to buy enough curiosity for the creative to do the explaining.
15. The workflow I’d reuse tomorrow
If I were making another launch film with Astra, I wouldn’t start with a prompt library. I’d start with a decision system.
Product truth. Write down what the product does, who it’s for, what’s genuinely different, and what cannot be misrepresented.
Creative tension. Find the part of the product that produces a feeling, a contradiction, or an uncomfortable question.
Cultural frame. Find the reference that compresses that feeling into something people already understand.
Idea stress test. Recognisable, tense, product-specific, novel, discussable, repeatable.
Message hierarchy. Decide what the viewer needs to know and in what order.
Script. Build the narrative before polishing the language.
Scene cards. Purpose, information, emotion, visual, transition, failure condition.
Reference system. Study references for mechanism, not aesthetics.
Production loop. Generate, review, diagnose, correct, regenerate, compare.
Approval gates. Idea, narrative, scene, visual, product, final cut.
Launch hypothesis. Why would anyone share, discuss, disagree with, or remember this?
That’s the system. The prompts are secondary.
16. The part I wouldn’t delegate
The real shift isn’t that AI made production faster. It’s that faster production changes the value of the decisions surrounding production.
When generating another version takes minutes instead of hours, making another version stops being the expensive thing. Knowing which version deserves to exist becomes the expensive thing. So the human role moves upstream. You decide what the product truth is, which tension is worth opening, which cultural reference carries enough weight, what counts as generic, what gets cut, and when the thing is finished.
Astra can sit inside every one of those decisions and make the loop dramatically faster. The creative thesis still has to come from somewhere.
For Moonshot the thesis was simple. An agent that knows enough about your life to act before you ask is useful precisely because it knows a lot about you, and that same fact can feel unsettling. So we stopped hiding the discomfort and used it as the opening line.
We built 1984’s Big Brother.
Astra helped us push on that idea, shape it, test it, produce around it, and keep iterating until it became a film. The 300K+ views came after. The more useful takeaway is the workflow that came before them.
Product truth → tension → idea → narrative → scene → generation → diagnosis → iteration → launch. The model made the loop faster. The job was still knowing what deserved to go through it.
