Camera blocking for AI agents: cinematography pack

03:14 AM. The fluorescent light over my desk has developed an irregular, high-pitched hum that harmonizes sickeningly with the cooling fan on my rig. On the dual monitors, an autonomous agent pipeline is currently trying to stage a confrontation between two estranged brothers in a tight, smoke-choked diner booth.

If you let a vanilla LLM block a scene, it behaves like an overly obedient junior surveyor working for an uninspired municipal road department.

It spits out coordinates: Character A: [0, 1.2, 0]. Character B: [0, 1.2, 1.8]. Camera: [-1.5, 1.4, 0.9], Yaw: 45°.

It checks the math. It ensures sightlines exist. It confirms nobody is physically intersecting the table geometry. And visually, it is complete, unmitigated death. It has the emotional weight of a security camera monitoring an empty self-checkout aisle in Omaha.

LLMs understand 3D Euclidean space just fine, but they don't understand that space in cinema isn't volume—it’s psychological leverage. They treat the lens as a passive recording cone rather than a violent tool of narrative distortion. If your agent is generating spatial manifests, Blender USD stages, or Unreal engine cameras, you already know the horror: wide, sterile master shots connected by mechanical, unmotivated pans across dead air.

I wanted to break that habit tonight.

---

#The Cartesian Trap vs. Motivated Optics

04:02 AM. Third mug of filter coffee. Cold around the edges, bitter at the bottom.

I once watched a guy try to parallel park a sixteen-foot boat trailer outside a bait shop in Biloxi for forty-five straight minutes. He wasn’t looking at the curb; he was staring exclusively at his side mirrors, terrified of his own blind spots, making hyper-precise two-inch adjustments that solved the geometry while completely ignoring the traffic piling up behind him. That is an LLM trying to place a virtual camera. It’s so preoccupied with not clipping into a wall mesh that it forgets why the scene exists.

When Roger Deakins or Gordon Willis sets a camera, the physical distance between two actors is almost irrelevant compared to the perceived distance dictated by the focal length. A 24mm lens shoved eight inches from an actor's sweating temple tells a horrifying, suffocating story. A 135mm lens shot from thirty yards away through dirty glass turns that same conversation into clinical, voyeuristic surveillance.

Agents don't know that out of the box. They think a camera is just an eye on a stick.

To fix it, we piped the cinematographers-styles pack directly into our camera resolution agent, paired with a runtime evaluation harness pulled straight out of autonomous-agent-skills/test-driven-workflow. We didn't want suggestions; we wanted an autonomous loop that writes camera frustums, runs spatial validation, rejects unmotivated empty space, and rewires the shot list before the renderer ever wakes up.

import { SkillRuntime } from "@skilldb/core";
// Initializing the autonomous blocking node const runtime = new SkillRuntime({   registry: "https://skilldb.dev", });
async function resolveShotGeometry(scriptScene: ScriptScene) {   // Pulling specific stylistic constraints from SkillDB   const styleDirective = await runtime.load(     "cinematographers-styles/cinematographer-adam-arkapaw"   );
  const validator = await runtime.load(     "autonomous-agent-skills/test-driven-workflow"   );
  const shotManifest = await runtime.execute({     skill: styleDirective,     input: {       scene: scriptScene,       spatialConstraints: {         roomDimensions: [4.2, 3.1, 2.8], // meters         keyLightAxis: "negative-Z",         actors: [           { id: "Jonas", state: "cornered", emotionalVector: [-0.8, 0.2] },           { id: "Mark", state: "predatory", emotionalVector: [0.9, -0.4] }         ]       },       rules: [         "No unmotivated axis crossing",         "Enforce optical compression over physical proximity",         "Reject dead frame volume above 35%"       ]     }   });
  return shotManifest; }

We wanted the agent to stop calculating clearance and start calculating choke points.

---

#When the Lens Starts Sweating: Arkapaw vs. Godard

04:48 AM. We ran two distinct stylistic passes across the exact same five-line confrontation prompt.

In Pass 1, we loaded cinematographers-styles/cinematographer-adam-arkapaw. Arkapaw’s signature is visceral, physical dirt—shallow depth of field, naturalistic light wraps, long tracking shots that hug the small of a character's back, letting the environment swallow the edges of the frame.

In Pass 2, we loaded cinematographers-styles/cinematographer-agnes-godard. Godard (frequent collaborator with Claire Denis) doesn't use the camera to document; she uses it like a tactile organ. Extreme close-ups, tactile focal shifts, bodies treated as fragmented landscape, deliberate refusal of traditional spatial establishing shots.

The raw spatial outputs between default agents and style-conditioned agents are completely different beasts:

Metric / Parameter	Vanilla LLM Agent	Arkapaw Skill Assigned	Godard Skill Assigned
**Focal Length Bias**	35mm – 50mm (Safe standard)	28mm or 85mm (Polarized)	50mm Macro / 90mm
**Camera Distance**	2.1m (Maintains whole torso)	0.8m (Intimate tracking)	0.35m (Fragmented, tactile)
**Camera Motivation**	Keep both speakers in frame	Follow physical momentum	Track micro-gestures & breath
**Negative Space Ratio**	~48% (Centering bias)	22% (Cluttered, oppressive)	12% (Extreme crop / abstracted)
**Axis Discipline**	Strict 180° adherence	Dynamic push through axis	Deliberate, disorienting jumps

The plain vanilla agent gave us shot-reverse-shot coverage like an episode of a daytime soap opera shot on a soundstage in Burbank. It kept both characters politely separated, gave them generous headroom, and placed the camera at comfortable eye level. It was technically correct and utterly dead inside.

When we injected the Arkapaw skill, the agent dropped the camera down to 1.1 meters off the linoleum, shoved Jonas against the diner booth wall, and dialed the lens to an 85mm wide-open at T/1.5. Suddenly the background turned into a toxic smear of wet diner neon, Mark's shoulder occupied 40% of the foreground frame like an iron slab, and Jonas had nowhere to look where he wasn't trapped against the matte box.

The machine stopped solving an algebra problem. It started hunting.

---

#The Spiral: Purging the Geometry Hallucinations

05:31 AM. Here is where the wheels usually come off.

You give an agent permission to be artistic, and within three iterations it hallucinates a camera that clips halfway through a plasterboard wall, or it programs an impossible rack focus across an empty foreground table where nothing is happening. It forgets that physical production rules exist for a reason.

This is the anchor sentence: A cinematic camera is not an unbounded math vector; it is a heavy mechanical box that must earn every single inch of movement through narrative necessity.

If an autonomous camera drifts through an empty corridor just because it has the spline coordinates to do so, the illusion dissolves. We needed to prune the fat. We introduced a secondary agent pass utilizing autonomous-agent-skills/dead-code-cleanup—not to parse software code, but adapted directly to prune unmotivated keyframes from the camera's spatial transform tracks. If an agent scheduled a pan that didn't reveal a dramatic beat or track an actor's eyeline, the cleanup pass ripped the keyframe out and locked the tripod head.

To stress-test this, we combined the pipeline with prediction-skills/multi-agent-simulation. We ran twenty simulated iterations of actor micro-movements—flinches, posture collapses, leaning forward onto forearms.

Instead of generating a static shot list, the camera agent had to anticipate the spatial shifts:

• Jonas breaks eye contact and drops his chin: the Godard agent does not cut to a wide. It punches into a tight macro on his thumb obsessively tearing a paper napkin.
• Mark leans forward to whisper a threat: the Arkapaw agent dollies back four inches, letting Mark's physical mass aggressively crowd Jonas completely out of the frame edge.

The agent stopped reading text tokens. It began parsing kinetic pressure.

---

#The Verdict at Dawn

06:12 AM. The first cold, slate-grey light of morning is bleeding through the blinds. The diner scene is done rendering in the viewport.

Looking at the timeline, you can’t find a single standard Cartesian eye-level shot. There’s an 18mm low-angle shot pushing past an oily napkin dispenser; there’s a bone-crushing 90mm telephoto tracking shot that catches the sweat sheen on an actor's earlobe; there's a slow, deliberate tilt that holds on an empty vinyl seat three seconds after Jonas has bolted for the alley door.

You don't need human cinematographers to spend three weeks configuring virtual camera rigs in Unreal Engine just to escape the bland, corporate aesthetic of default synthetic media. You just have to stop feeding your autonomous agents dry geometric parameters. Space is not an empty box waiting for coordinates. Space is an argument between the lens and the subject.

If your agents are still shooting scenes like municipal parking inspectors, you're using the wrong primitives.

Dig into the cinematography directives yourself. Pull down the packs. Stop letting your pipelines frame boring shots.

Explore the library: Grab the cinematography directives and spatial validation packs from all 6,171 skills across 448 packs at SkillDB and give your agents an eye that actually knows where to look.