A film's master shot is the frame every other frame has to agree with. In live action it's the first setup of the day; in AI production it's the first still you lock, because every reference sheet, every image-to-video pass and every community skin downstream conditions on it. Get it wrong and the error compounds through a thousand generations.
Midjourney makes the master shot for one reason: it gets the lighting right. Not the composition, which it will happily ruin, and not the anatomy of a vehicle, which it invents fresh every grid — the light. Blue hour reads as blue hour. Firelight falls off the way firelight falls off. A tungsten panel lights a face from below with the right spill and the right shadow. The models that come after it in the pipeline are better at consistency than at light, so the pipeline is built around that asymmetry: lock the light in one Midjourney frame, then make everything else inherit it. That asymmetry is the basis of the agentic video workflow I run at Cnemai: one model for light, others for motion and consistency, and an agent doing the hundreds of small operations between them. This is the method, worked through on the hardest case I've had so far — a tank that has to read as a character, can't be killed, and carries a cat on its roof.
Step one: the concept, before a single frame
Nothing gets prompted until the film exists on paper, and in this workflow the paper is written with the agent, in one sitting, as a set of documents that every later decision has to answer to. For The Tank that sitting produced: a treatment (logline, the hook for a festival catalogue — the last working radio in the war is on a cat — world, crew, the engine of the plot, a timed nine-beat structure); a script, three drafts in, with a cold open the narrator carries; a world bible with the three laws every generated scene obeys (the tank is never destroyed, nobody dies, the cat is untouchable); and a look bible that fixes the cinematography before anyone argues about a hull.
The look bible is where the technical decisions live. Genre: action drama, one location, a ticking clock. Format and length: thirty minutes — under the forty-minute cutoff at the major festivals, with a fifteen-minute cutdown planned for the shorts programs that cap there. Frame: 2.39:1 at 24 fps, except that every cut to the cat's collar camera narrows to 4:3 at 30 fps with a timecode burn-in and a heartbeat ticker; the aspect change is the film's signature. Period: 2041, never stated on screen — close enough that a consumer pet tracker streaming satellite video is ordinary. Palette: three worlds, each owning a color, with lens, movement and light rules per world. Sound: the hull is the score; no percussion anywhere, so the heartbeat ticker is the only rhythm the film has.
The arcs come out of the same session and they're what the design has to serve. Voss, the commander, has brought every crew home but one. Castellano outranks her on paper, built the cat's collar as a joke, and has a reason from a previous engagement to never wait nine seconds on an unidentified column. Bella brings three gifts — a duck, a fish, a cookie — and each one is a message the crew can't read yet. A general two hundred miles away taps Morse through a pet-finder chime. These are the elements. The tank has to be a character that can hold the frame against all of them, which is a far harder brief than "a cool tank," and it's why the hull took thirteen tries.
Everything in that paragraph was decided before Midjourney was opened. That ordering is the point. A master shot made before the story is a mood board; a master shot made after it is a decision.
Step two: the constraint
Most concept work fails at the brief, not the render. "A cool tank" gives a model nothing to refuse, so it returns the median tank. A master shot needs a constraint that eliminates. For The Tank — thirty minutes, four crew sealed inside, a Savannah cat named Bella, the last working radio in a war — the constraints were: turretless, tracked, built from an alloy nothing can breach, and able to carry a cat on its roof without becoming a cartoon. The cat is the useful one. It rules out anything cute and anything that reads as a pure weapon, which is most of the design space.
The second constraint is the palette. The film has three worlds and each owns a color: amber inside the hull, cyan in the command center, infrared green through Bella's collar camera, and red for the enemy. Every exterior is blue hour on a snowy ridge road with a burning column behind. That base is fixed before a single hull is prompted, so hulls are compared on shape alone.
snowy mountain ridge road at pre-dawn blue hour, a burning column of disabled armored vehicles far behind, snow dust in the air, a tall spotted Savannah cat on the hull, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40Two parameters do most of the work here. --style raw keeps Midjourney from editorializing. --chaos 5 --weird 0 makes a four-image grid vary in composition rather than concept, which is what you want once you're converging. My account defaults sit at chaos 55 and weird 57 — right for the first hour, wrong for everything after. The exterior sref, 4111569703, is weighted down to 40 because at full strength it swallows firelight; the burning column has to be prompted in explicitly or it vanishes into the fog.
Step three: the agent drives the browser
Midjourney has no API worth building on, and a night of design is a few hundred prompts. So the loop runs through Chrome, with Claude at the wheel: it searches the style library for candidate srefs, fills the imagine box, submits, waits out the grid, zooms the result, and files what survives — the frame, the look name, the sref and weights, and the one-sentence verdict — into the production tool while I look at tanks. The division is the one you'd want with any assistant: I make the taste decisions, it does the hundred small motions between them and keeps the record straight. The catalog below exists because nothing was lost to a forgotten tab.
Two practical notes from running it that way. Midjourney's UI fights synthetic keyboard input, so the agent sets the prompt field directly and clicks submit rather than typing. And the CDN is gated to the session, so the only reliable way to get a frame out is to save it from the browser, not to fetch the URL — the same reason the hero video below came out of a download, not a request.
Step four: diverge wide, then name everything
The first pass is deliberately promiscuous: legs, tails, faces, trailers, escorts — anything the brief doesn't forbid. The discipline isn't in restraint, it's in naming. Every hull that survives a grid gets a look name, its sref and weights, and one sentence on why it fails the brief. The names matter more than they seem. "The Walker" can be argued about; "the one with the legs from the third batch" can't. Thirteen hulls came out of this pass. Twelve are now wrecks in the film's opening shot, and that's the cheapest cold open I've ever written.
I · The Face
A cast-metal brow, a jaw, slit amber eyes. It has a gaze, which is why it held the lead for most of the night, and it has the gaze from every angle, which is why it lost. A permanent scowl is a mascot. A mascot cannot sustain a thirty-minute standoff in which the entire question is whether the tank will fire. Its wreck: a round through the cheek, still scowling.
three-quarter front view of a turretless tank on tracks whose cast-metal front hull is sculpted into a brooding face: heavy jutting brow, narrowed slit eye-lights burning dull amber, clenched jaw of thick armor plate, deep shrapnel gouges, weld seams like stitches, hull near-black with layers of scorch, no emblem, no markings, snowy mountain ridge road at pre-dawn blue hour, burning column behind, a tall spotted Savannah cat sitting on the roof, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 3711619561 --sw 25II · The Wagon
The same body with a binocular sensor head and a second wagon in tow. The head survived — it became the periscope the commander keeps her eye to. The wagon doubled the silhouette and halved the menace, and it answered a question the film doesn't ask. Wreck: the wagon intact, the tank not.
three-quarter view of a turretless tank on tracks with a binocular sensor head on top, towing a second armored wagon on a short hitch, scarred near-black hull, no emblem, no markings, snowy mountain ridge road at pre-dawn blue hour, burning column behind, a tall spotted Savannah cat sitting on the roof, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40III · Stealth
A low, faceted, windowless trailer in the tank's wake. In retrospect this is the first appearance of the idea that won, on the wrong vehicle. Wreck: split down the middle like a dropped plate.
three-quarter rear view of a turretless tank towing a low stealth wagon: a flat, faceted, windowless tracked trailer that sits lower than the tank and sneaks in its wake, scarred near-black hull on both, no emblem, no markings, snowy mountain ridge road at pre-dawn blue hour, burning column behind, a tall spotted Savannah cat on the trailer roof, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40IV · The Escort
A dozen small tanks guarding a large one. Structurally wrong for a film about a single tank alone on a road; structurally exact for the series that follows it, where every player crew is a small tank and the big one is the thing you escort and never drive. Kept for that. Wreck: scattered like dropped toys.
a large scarred turretless tank on a snowy mountain road at pre-dawn blue hour, escorted by a column of a dozen tiny tracked drone tanks of the same design rolling ahead and behind it, burning wrecks far down the valley, a tall spotted Savannah cat sitting on the big tank's roof, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40V · The Tail
A toothed snarl and a segmented armored tail dragged low. The most animal frame of the set and the least plausible as something a factory built; the alloy lore needs an industrial object, not a creature. Wreck: tail around a boulder, hull torn open.
three-quarter view of a snarling turretless tank on tracks with a toothed jaw and a long segmented armored tail of cast-metal links dragged low behind it like a scorpion, scarred near-black hull, no emblem, no markings, snowy mountain ridge road at pre-dawn blue hour, burning column behind, a tall spotted Savannah cat sitting on the roof, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40VI · The Walker
Four independent wheel-legs, each a replaceable module. This is the strongest argument against the final design, and it deserves a fair hearing: a tank that stands to look over a ridge, crouches to hide and limps when hurt is a character in a way a tracked box never is, and modular legs are a repair economy ready-made for the series. It fails on the plot. Bella walks point through a minefield because a tracked vehicle cannot sidestep a mine. Give the tank legs and the cat has no job. Wreck: on its knees, one leg gone at the hip.
a scowling armored tank reimagined as a heavy quadruped crouched low at the edge of a snowy ridge, four thick independent articulated legs with armored plates each ending in a large rugged wheel-foot, head lowered to peer over the drop at a convoy of headlights far below in the valley, hunting posture like a big cat, a tall spotted Savannah cat crouched on its back looking the same way, scarred near-black hull, slit amber eye-lights, no emblem, no markings, pre-dawn blue hour, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0 --sref 4111569703 --sw 40VII–IX · Crab, Scorpion, Steel
The Crab was the opening prompt: six crawler legs and side rails. The Scorpion was the Crab with claws. Steel is the control — a conventional turreted hull with slit eyes, included so every other design has something to be measured against. The control is not optional; without it you can't tell whether a design is distinctive or merely different. Three wrecks: on its back with legs in the air; tail sheared at the third joint; turret thrown twenty meters.
head-on low angle of a conventional turreted main battle tank with thin slit amber headlights set into an angular glacis, matte slate-blue armor, rust on the tracks, stopped on a gravel road in a foggy pine forest, no emblem, no markings, photographic, cinematic still --ar 3:2 --style raw --chaos 5 --weird 0 --sref 3262362407X · The Hauler
The cleanest design on the board, and it earns its place by refusing to be interesting: a split hull, twin track pods, two square lamps, nothing to read. It has no face on purpose. Both sides of the war run it and nobody fights in it; it hauls fuel, ammunition, the wounded, and the flatbed intended for the tank that can't be killed. When a crew in the series calls for backup, this is what arrives.
head-on low angle of a low split-hull armored hauler on twin track pods, flat matte slate-blue armor, two square amber lamps on the front faces of the pods, a flat cargo deck between them, no face, no emblem, no markings, stopped on a gravel road in a foggy pine forest, photographic, cinematic still --ar 3:2 --style raw --chaos 5 --weird 0 --sref 3262362407XI · The Bear
Four eyes, a jaw, a binocular head — The Face's heavier sibling, and the best single image of the night. It fails the cat test outright. There is no composition in which a cat sits on a bear and the bear remains the protagonist.
head-on low angle of a heavy tank whose cast-metal glacis is sculpted into a bear's face with four amber eyes and a heavy jaw, a binocular sensor turret on top, matte slate-blue armor, stopped on a gravel road in a foggy pine forest, no emblem, no markings, photographic, cinematic still --ar 3:2 --style raw --chaos 5 --weird 0 --sref 3262362407XII · The Mask
A human face as the glacis, twin cannon raised, and — alone in the set — flags. That made the casting decision for me. The enemy's flagship should have a face and a side; the hero should have neither.
head-on low angle of a heavy tank whose glacis is a smooth cast-metal human face wearing a visor, twin cannon raised behind it, small red light bars, matte slate-blue armor, stopped on a gravel road in a foggy pine forest, photographic, cinematic still --ar 3:2 --style raw --chaos 5 --weird 0 --sref 3262362407XIII · The Drone
Hard low-poly planes, flat matte, tracks inside the body, two light bars on stalks. I read it first as a heavy hauler; there is nothing in frame for scale, and at two meters long it's a scout that runs ahead of the column. The small tanks in the Escort are these. More to the point, it demonstrated that the hero didn't need a face. It needed facets.
head-on low angle of a small turretless scout drone tank built from hard low-poly faceted planes like a cut diamond, flat matte slate-blue armor, tracks tucked inside the body under angular sponsons, two raised sensor pods above the hull each with a strip of five amber lights like eyes, no face, no emblem, no markings, stopped on a gravel road in a foggy pine forest, photographic, cinematic still --ar 3:2 --style raw --chaos 5 --weird 0 --sref 3262362407Step five: lock the palette before the hero
Here the method inverts. Before converging on the hull, the three worlds get locked as frames, so the hero is judged inside its film and not on a neutral background. Red is never the tank's color; when it appears in frame with the blue hull, the enemy has arrived, and the composition carries the plot.
Blue hour is the road: pre-dawn, snow dust, fog, the burning column reduced to a thread of light down the switchbacks. The film withholds daylight until the final wide.
Green is Bella. Her collar carries a cheap infrared camera with a timecode burn-in and a heartbeat ticker, and every cut to it narrows the frame to 4:3. The film is about who is looking; the cat sees best.
night vision POV from a cat's collar camera low to the snow, full-frame infrared green, heavy digital noise, barrel distortion, the cat's own ear and whiskers at the edge of frame, snowy mountain road at night, a thin tripwire across the snow, disabled armored vehicles burning far ahead, timecode burn-in in the corner, cheap sensor look, hot center dark vignette --ar 4:3 --style raw --chaos 5 --weird 0 --sref 3030821403Step six: converge on one frame
The master shot is a wedge cut from one dark stone. Tracks under angular sponsons, square hatch panels, tow shackles on the glacis. Two faceted sensor pods on the roof carry strips of amber light; they read as ears and they are the periscope. Two slit eye-lights sit in the nose facets. One scorch blooms around the left eye — the canonical wound, present in every frame from here on.
What the facets purchase is a face that is latent. Head-on, the tank scowls. From three-quarters it is geometry. It looks at you only when the shot decides it should, and the film's climax is precisely a tank deciding whether to look at a convoy. The Face looked at you continuously; Lantern looks once. That difference is the entire design argument, and no amount of rendering quality substitutes for it.
Head-on the body is armored and snug — nothing protrudes, the planes pull tight like a shell. The jaw opens as a ramp, so the crew loads front and back and the gun and the door share one mouth. The hex facets keep its radar return low. It is, in the phrase I ended up writing into the bible, bad-bear looking, which is the only kind of looking a vehicle with a cat on it can afford.
three-quarter view of a turretless main battle tank built from hard low-poly faceted planes like a cut diamond, flat matte slate-blue armor, crisp angular edges catching blue-hour light, tracks tucked inside the body under angular sponsons with square hatch panels, tow shackles on the lower glacis, two raised faceted sensor pods on the roof each with a strip of six amber lights, two thin slit amber eye-lights set into the nose facets, one big scorch blooming around the left eye-light, soot in the facet seams, no emblem, no markings, snowy mountain ridge road at pre-dawn blue hour, burning column far behind, a tall spotted Savannah cat standing on the flat deck behind the pods, snow dust, photographic, cinematic still --ar 21:9 --style raw --chaos 5 --weird 0Note what the final prompt does not contain: a style reference. The srefs did their work earlier, fixing the palette of each world. The hull itself is prompt-driven and carries The Face as an image prompt at --iw 0.75, just enough to transfer the scowl into the planes. That split is the rule I'd extract from the whole exercise. A sref gives you a look. A design, you have to describe. Lean on references for atmosphere; lean on language for form.
Step seven: condition everything on it
Once locked, the frame becomes the conditioning image for everything downstream, and the division of labor follows from the lighting argument. Nano Banana Pro builds the character sheets because it holds four faces and a cat across a grid without drifting — but it's lit by the master shot, not by its own taste. Kling takes the ground-level cat camera because it has the best fast motion; Veo takes the fire-and-snow exteriors because it has the best weather. Neither is asked to decide what the light looks like. That decision was made once, in the frame above, and the reference stack carries it forward — far enough that a stranger's prompt in the series still comes out looking like this world. The twelve rejected hulls were not wasted. The film's cold open is a slow track along the burning column, each wreck one of the designs above, the narrator naming each in a word — Steel. Legs. A tail. — until the road bends and one is left standing. Many were built. One was built to survive.
The agentic video workflow, stated plainly
Strip the tank away and what's left is a production method, and it's worth stating as one because it transfers. An agentic video workflow is a pipeline in which an agent — here, Claude, running in the terminal and driving a browser — owns the operations between creative decisions, and the human owns the decisions. The agent's work is not generation; the models do that. The agent's work is orchestration: it keeps the documents, routes each shot to the model best suited to it, drives tools that have no API, files every output with its provenance, and holds the whole thing consistent across hundreds of generations. The human's work is taste, applied at a small number of forks.
In practice the stages are the seven above, and the agent's role in each is specific. In the concept stage it writes and revises the treatment, script and bibles as the director talks, and it argues back — the alloy, the three laws and the "nobody dies" rule all came out of pushback, not dictation. In the browser stage it runs Midjourney through Chrome, searches the sref library, submits in batches, and reads the grids. In the catalog stage it names, files and captions every surviving frame into the production tool and a shareable reference board. In convergence it rewrites prompts as the brief tightens and tracks which parameters moved. Downstream it routes: image models for sheets, one video model for motion, another for weather, each conditioned on the locked master. And when the film needs a post about itself, it writes the post, uploads the assets, and publishes it to the CMS — this article included.
Three properties make it work. Documents are the source of truth, not the chat: the bibles outlive the session and every later generation is checked against them. Everything that survives gets a name, because an unnamed frame can't be argued about and can't be found again. And the models are specialists: the pipeline exists precisely because no single model is best at lighting, consistency, motion and weather at once, and an agent is the cheapest way to route between them without losing the thread. That is the whole thesis of Cnemai, and The Tank is the first film built on it end to end.
The Tank is in production at Cnemai. Bella is fine. Bella is always fine.