action_man
Open ↗Browser-based action prototyping playground.
Writing, experiments, and field notes from Joel Citron — building enterprise AI systems by day, prototyping at the edge of the model after hours.
Began as a creative director in 2005, pivoted to development in 2008, and have spent the last decade-plus marrying the two — building enterprise attribution systems at AI Media Group, where I'm now CTIO.
The thesis hasn't changed: a design-trained engineer makes better technical decisions because they think about the human standing in front of the system, and a engineering-trained designer ships things the design-only crowd never finishes. acidlemon is where I write about both sides.
2D samurai brawler in a late-Edo bloody-woodblock style. Every sprite generated through a local ComfyUI + Flux art pipeline. Proof of concept.
Head-tracked rail shooter, a Space Harrier homage. The webcam re-projects the scene with off-axis projection, turning your monitor into a window. Original geometry and audio.
Three.js maze runner with post-processing shaders, outline pass, and effect composer.
First-person maze, simpler engine. Find the exit.
Three.js first-person shooter prototype. Mouse + keyboard.
Side-scrolling platformer, vanilla canvas. Arrow keys + space.
Browser-based action prototyping playground.
Personal real-estate tracking + analysis tool for the NJ market.
Operator-persona executive site — Astro + Tailwind, Metric Stage design system.
This site — Calibrated Gantry design system, bento grid, dark mode.
A woodblock samurai brawler and a head-tracked rail shooter, both built as proof of concepts. The point wasn't the games. It was being a beginner again, where the agent can't fake it.
The 2026 AI cost reckoning trains your attention on the token meter. When I traced where an agent's spend actually went, the model was the easy part. The leak was the warehouse the agent queries.
Karpathy joined Anthropic to run autoresearch at frontier scale. The same closed loop is the operator-tier curriculum now, one tier down.
Three failure modes in agents that synthesize work you own, and why the third one breaks the whole review chain.