On this page

Motivation#

Every AI engineer I know is building some version of the same thing right now: a personal software factory, where agents write more code than any human could possibly review. At some point it stopped feeling like a workflow and started feeling like a creed.

And no, I don’t want to see your software factory. 😂

So I recast the most famous creed in film: the barracks scene from Full Metal Jacket (1987), where the recruits recite their creed to their rifles before lights out. In my version the rifle is your agent, the recruits are the AI engineers I interact with the most on X, and the drill instructor is Jensen Huang, the man who coined “AI factories”, in a leather jacket instead of a uniform.

Watch original video
The original: the barracks creed scene from Full Metal Jacket (PG version).

Matt Pocock delivers the hook, Peter Steinberger holds the line in the wides, swyx (who coined “AI engineer”) gets to chant it in his own close-up, Simon Willison stands fire watch and gets the last word, and poteto’s avatar hangs on the office wall as the “PRs” poster.

How it was made#

The whole thing came out of one long collaboration with Claude Code (Opus 5.5): it wrote the scripts and review pages, and I reviewed every round on them and sent back notes.

  • Script: every parody line is locked to the original line’s start and end, within two syllables, so the new words land on the same beats as the film.
  • Voices: the recruits’ creed is a single sung take from Eleven Music, cut at its pauses, so the crowd sounds like one platoon. Jensen’s orders are a Fish Audio clone of his voice, built from ten minutes of clean interview audio. The fire watch is an ElevenLabs designed voice.
  • Shots: every shot is regenerated: Nano Banana Pro keyframes built from character sheets of real headshots, animated with H3 Max with each line’s audio pinned, then lip-synced again wherever the mouth didn’t sell it.
  • Title, captions, and credits: rendered locally in HyperFrames, with every caption word force-aligned to the vocals and lit as it’s sung.

I also tried a second direction, lip-syncing the cast onto the original footage, and dropped it: the fully generative cut is the project.

Matt’s character sheet: real headshots in, consistent likeness out.

Key decisions and lessons#

  • The McGurk effect: or what you see changes what you hear. Jensen’s opening “PROMPT!” kept sounding like “PRONT!” even though the isolated audio was perfect. It turns out this was because the generated mouth never closed for the M and P. This was so surprising to me that I created a write and explainer video here. The fix was a new lip sync over the exact same audio.
  • Pin both ends of a shot. Every H3 prompt pushed the camera in until I anchored each shot’s last frame to its keyframe too. H3 also renders its frames about 2.5% wider than the keyframe it starts on, which snapped visibly in the first few frames; exact 16:9 keyframes and holding the first and last frames fixed it.
  • Small faces need help. The fire watch’s face is about 100 pixels tall under a helmet, and two lip sync models passed it through untouched. Lip-syncing an upscaled crop of the shot, then blending back only the mouth and jaw, finally got him to say his line.
  • Never re-time generated audio. Stretching a read to fit always sounds worse than regenerating it, so every take is placed whole.
  • Measure the captions, don’t eyeball them. Whisper’s word times run early after pauses, so the captions are force-aligned against the separated vocals and checked against the finished mix: line starts land within 25 ms of the voice.
Captions in the video’s own style: each word lights up gold as it’s sung. Scene featuring swyx.

Cost#

A 76-second, fully AI-made parody with a title card, word-timed captions, and end credits for the cast. It took 117 fal generations (about $97) and a lot of review rounds.

Tools#

Credits#