Skip to main content
Michael Poynor
All case studies

Case study · AI-assisted production

Directing AI in Blender

I ran three rounds of short films with two AI coding agents, Claude Code and OpenAI's Codex. They built everything in Blender with code. After each round I watched the films and gave notes, and my notes became the rules for the next round. Round 3 ended with this launch film for a camera that doesn't exist.

PLEAT 01 · "Unfold the Moment" · 0:30 · 1920×1080 · 24 fps · original score at 120 BPM · made with Claude Code
My role
Concept, briefs, production design, notes, final review
Made with
Claude Code, OpenAI Codex, Blender 4.5, Python, FFmpeg
Output
Six films over three rounds, plus a full launch kit for PLEAT 01
Context
Self-directed experiment. All products are fictional.

The setup

I wrote the briefs and gave the notes

I wanted to see if I could direct AI the same way I work with a post team, with a clear brief up front and notes on every cut. Each round had one brief with a broad challenge. Claude Code and Codex each came up with their own concept and built it in Blender with Python, including the models, lights, animation, and sound.

Between rounds I watched the films like any other edit and wrote notes. For Round 3 I planned the production with AI help. I set up the departments, the handoffs between them, five review checkpoints, and the standards for each one. An AI director ran it from there, and I reviewed the finished films.

The films

Round 1

Product study

AXIOM Claude Code · 7.5 s · silent

Gold, steel, and black parts lock together into a puzzle on a marble plinth in an empty studio.

VESPER Codex · 7 s · silent

Ceramic and titanium arches settle around an amber optic.

Round 2

Motion piece

MIDWAY Claude Code · 10 s · original mix

A pop-up book bursts into a glowing paper carnival in a child's bedroom at night.

SUNBREAK Codex · 8 s · original mix

A mechanical seed winds up, leaps across a meadow, and blooms into a flower.

Round 3

Product launch

PLEAT 01 Claude Code · 15 s cutdown

A pocket-folding instant camera. The :15 got its own music edit instead of a trimmed :30 track. The full :30 is at the top of the page.

NYRALO / FIELD 01 Codex · 24 s · original mix

A portable directional light. Turn the dimmer, pivot the head, and the light reveals a coastline at dusk.

Both agents got the same brief each round and came up with different ideas. I ran this as a creative experiment, so the pairs aren't a software benchmark.

The feedback loop

My notes became rules

After Round 2 I gave the kind of notes I'd give in any review:

After the camera moves closer to the Ferris wheel and dollies across it, the camera kind of stutters. It should be one seamless motion, easing into each moment.
The fireworks stop and hang in the air for a beat before popping. Keep them moving up until they pop, never losing energy.
Can you build out the room a bit more? A window showing it's night time, and a street light shining in.

Those shots got fixed in MIDWAY. Then each note went into the Round 3 brief as a standard every shot had to pass.

  • "It stutters" became a camera rule. Each move is one path with one speed curve, and the speed can't jump or dip from one frame to the next. Every PLEAT shot was measured against it.
  • "Never losing energy" became a physics rule. Moving parts arrive with speed and hand it off through a click, a bounce, or a slide, so nothing just hangs there.
  • "Build out the room" became a design rule. Every scene needs a real setting, so PLEAT has three: a light-painted stage, color backdrops, and a sunlit windowsill.

Codex got notes too. I liked the SUNBREAK character but asked for a more natural world and a better sky. Codex rebuilt the meadow and the sky and left the character and animation I'd already approved alone.

Plot of camera speed per frame for the PLEAT unfold shot: one smooth rise and a long, soft landing.
Camera speed per frame in PLEAT's unfold shot, checked against the stutter rule. It rises once and lands slowly, with no jumps.

Round 3

Nine departments and five checkpoints

For PLEAT I set the production up like a studio job. There were nine departments: research, brand, industrial design, modeling, lighting, animation, edit, sound, and delivery. Each one owned its own files. Work passed between them through shared specs, like the product's dimensions, the logo geometry, a rig definition, and a beat grid with every sync point in the film down to the frame. Nothing moved forward until it passed a checkpoint.

  1. G0ConceptName, logo, spec, and research agree.
  2. G1ModelThe mechanism works and the details read up close.
  3. G2AnimaticMotion and timing locked to the music in gray.
  4. G3LookLighting approved shot by shot within the render budget.
  5. G4DeliveryEvery file checked before it ships.

At the animatic checkpoint, the exploded view of the camera was correct but timid. It went back and got staged bigger before any final render time was spent on it.

The PLEAT production team

One Claude session was the director. It briefed up to four specialist agents at a time, reviewed their images, and made creative calls within the standards I'd set.

DepartmentWhat it delivered
ResearchMarket report with dated, cited sources, a 15-product competitor matrix, and a price ladder
BrandName check, logo exploration, custom wordmark, color and type system, brand book
Industrial designProduct spec, mechanism and clearance notes, dimensioned drawings, explode plan
ModelingA 71-part rigged model built from the spec, with collision checks across the mechanism
LightingMaterials, three sets, and one color look for the whole film
Animation15 shots in two teams, every hit on its frame of the beat grid
EditStoryboards, edit lists for the :30 and :15, a cut list with a reason for every cut
SoundOriginal score and effects synthesized in code, with separate :30 and :15 mixes
DeliveryEncodes, 4K stills, contact sheet, and a full technical check

Calls the AI director made

  • The brief asked for a 108 mm print to come out of a 96 mm-tall camera, which can't work. The director caught it, switched to a landscape print, and had the body reworked around it.
  • Its first logo, three chevrons in a rounded square, read as a fast-forward button, so it threw that out. The final mark is the bellows seen head-on with a hexagonal iris. It stayed away from a circle on purpose, because a circle in a rounded square reads as Instagram.
  • Research recommended the cheaper film format. The director went with the wider one so the developing print in the last shot could show the Round 2 carnival frame without a heavy crop.
  • The first encode was missing 12 frames because of a bit-depth mismatch on the logo layer. The final check caught it, and the files were re-encoded before they got to me.
Rejected chevron logo explorations.
Rejected: chevrons that read as a media control.
The final PLEAT logo system.
Approved: the bellows seen head-on.

Process gallery

A gray-box version of the camera from the first hour.
Hour one: a gray stand-in so departments could start in parallel.
The finished camera unfolded on the black stage.
The same shot in the final film.
Storyboards for all 15 shots.
Storyboards for all 15 shots, with camera moves, key frames, and sound cues.
The animatic at checkpoint G2: quick preview renders with the shot, frame, and beat burned in, cut to the real score.
Clay renders of the camera closed, open and exploded.
The model closed, open, and exploded.
Price ladder of instant cameras.
Research: where a $449 folding camera would sit. A positioning hypothesis for a fictional product.
Three frames of the print developing into the Round 2 carnival.
The print develops into the Round 2 carnival frame.
Contact sheet of all 15 PLEAT shots.
All 15 shots of the final film.

The numbers, and what they measure

MeasurePLEAT result
Final shot frames784 frames, 5.25 s per frame on average, 68.7 GPU-minutes on an RTX 4080 SUPER. Previews, look tests, and revisions are not included.
Camera smoothnessAll 17 shot files passed the stutter check (speed can't change more than 12% between frames, with a tolerance for very slow moves).
SoundBoth mixes at −16 LUFS integrated and −1.5 dBTP. Every effect starts within 5 ms of its picture frame.
DeliveryFrame counts, durations, color tags, decode accuracy, and file integrity all checked. Anything that failed was fixed before delivery.
Elapsed timeAbout 21 hours from kickoff to delivery. Most of that was waiting for my AI plan's usage limit to reset, which happened four times.

These numbers tell me the camera doesn't stutter and the files are correct. Whether the film is any good, I decided by watching it.

What didn't work

  • The agents can't really watch or listen. On PLEAT they reviewed sampled frames, contact sheets, motion plots, and audio measurements. They never saw it play in real time or heard the mix. The Codex productions checked sampled moments during actual playback. After the rounds I had Claude do a full sound pass on MIDWAY. Its measurements all passed, and I still heard static that it couldn't.
  • The work stopped four times when my plan's usage limit ran out. Nothing was lost, since every agent picked up where it left off. Next time I'd plan narrower tasks, run fewer agents at once, and use checkpoints that survive an interruption.
  • One animatic briefly mixed new frames with stale ones from an earlier version. Every rendered frame needs to carry the version it came from.
  • PLEAT still has flaws. The on-screen UI is too small to read, and a few internal parts pass through each other for a frame or two in the exploded view.

At work

Where I'd use this on a real production

Previz first. I could send a brief in the evening and review an animatic cut to temp music the next morning, so motion and timing are locked before anyone starts lighting.

Versions next: cutdowns, colorways, and aspect ratios built as variations of one project, with every cut still on the beat. I'd also have it run technical QC on every export, checking frame counts, loudness, color tags, and dropped frames, so my review time goes to the creative. And for product work, if the model is built from a spec file, an engineering change means a rebuild instead of a remodel.

I'd keep the story, the notes, the final cut, and the client conversation.

If you try this

What I'd tell another editor

  • Write the brief like a contract, with durations, frame counts, formats, and who owns which files.
  • Give notes that can be checked. "It stutters" stayed fixed once it became a rule with a number on it.
  • Lock motion before look. An animatic is still the cheapest place to change your mind.
  • Lock what you've approved, like the SUNBREAK character, so revisions can't touch it.

Credits

Concept, briefs, production design, notes, and final review: Michael Poynor. AXIOM, MIDWAY, and PLEAT 01 were made with Claude Code (Anthropic). VESPER, SUNBREAK, and NYRALO / FIELD 01 were made with OpenAI Codex. All six were built in Blender with Python and FFmpeg and rendered on one workstation.

Source credits: PLEAT 01 uses the Manrope typeface (SIL Open Font License). Its geometry, textures, music, and effects were generated in code, apart from the MIDWAY frame on the final print. SUNBREAK uses five CC0 assets from Poly Haven. NYRALO uses a CC0 stone texture and Source Sans 3 (SIL Open Font License). All products are fictional, and pleat.camera is not a real website.