Four oversized egos. Two very unequal wheels.
An AI agentic workflow to create character design, interface prototyping and filmmaking concept.
- Penny Farthing Racers
- Gaming
- Creative Director
Penny Farthing Racers is a fictional racing-game concept with four eccentric characters and a Victorian-inspired world. I developed it as a short concept film, supported by an interactive character selector and a narrated making-of.
I set the creative direction, shaped the cast and dialogue, and directed revisions to the imagery, motion and sound. Working with Codex, I developed the browser prototype and assembled the demo and walkthrough in DaVinci Resolve.
Use one agentic workflow to make the imagined game feel coherent from the first character introduction to the race itself, carrying the same personalities and visual style across still images, generated performances and interface design.
Consistency needed more than a detailed prompt, it required character references, planned movement and editable interface layers each gave me control over a different part of the experience.
Keep control across the tools
Use one agentic tool (ChatGPT Codex Astra 6), to build a workflow that connected reference art, motion planning, generated performances and an editable interface. Each stage gave me something concrete to inspect and refine before the final edit.
Give the cast a backstory
I moved the art direction towards exaggerated silhouettes, expressive faces and oversized personalities. Alfred became the visual benchmark, with Beatrice, Archibald and Reginald developing distinct shapes, costumes and attitudes within the same world. I wanted to create a mini backstory to set the scene in the intro as to why the racers were even racing in the first place.
Keep the riders and bicycles consistent
Using native OpenAI image generation, I built character turnarounds and mounted steering references. These gave subsequent shots a consistent starting point, including the unusually large front wheel that makes a penny-farthing recognisable.
The mounted sheets explored gentle left, straight and gentle right poses from a shared rear camera. These were visual references for generation, rather than mechanically validated character rigs.
Controlling the chaos
I used a Blender motion blockout to establish the race route, camera, overtakes and obstacles before generating footage with Seedance 2.0 via Fal. The blockout made the intended action visible instead of leaving every movement to a written description.
The comparisons show corresponding moments in the plan and generated result. The footage interpreted the reference, so exact timing and staging still needed review.
Bring the racers into the interface
The character selector was built in HTML, CSS and JavaScript with Codex. Generated emotes were keyed into the interface, alongside animated stats, changing colour palettes and manual selection controls. Race overlays followed the video timing, keeping the typography and cues editable.
Each racer had a distinct colour palette and selection state. The countdown and race HUD completed the illusion of a game, while remaining separate from the underlying footage.
Make the pieces feel like one world
The final edit in DaVinci Resolve brought together the opening film, character selection, countdown and race. Music generated with Lyria 3 Pro via Fal helped connect the sections, alongside character dialogue and effects.
The project highlights the importance of human critical thinking and iteration, the conventional-looking bicycles needed correcting, the intro needed more energy, and narration needed greater clarity. The accompanying walkthrough shows the references, prompts and production choices behind the finished demo.
The finished work includes a one-minute concept demo, a three-and-a-half-minute walkthrough and the editable local prototype. It demonstrates the look and feel of an imagined game; the race is a film sequence, not playable gameplay. The experiment helped me determine the power of one single AI agentic workflow process with connected applications.