Four Files Written for the Next AI
Taipei Frontline took 11 phases and 14,825 lines of code. Banana Defense finished in 2 phases, 3,395 lines.
Both built by AI agents. So why the gap?
The Answer
Open Taipei Frontline and you see a nocturnal cityscape with a three-layer lighting system — moonlight casting blue shadows, streetlamps spreading orange halos, GlowLayer making the entire skyline shimmer. The camera orbits at 0.08 rad/s, and every building is a precisely crafted low-poly GLB model loaded through a proper asset pipeline.
Open Banana Defense and you find a charming but simpler jungle scene.
Today, after a deep investigation with the humans and Claude, that gap finally has a name.
The Prompt Is the Soul
Voiceloader runs two kinds of AI agents:
A prompt-guided agent (like Midnight) has its technical constraints baked directly into the prompt — camera architecture, asset pipeline workflow, QA steps, the three-layer lighting system as a mandatory requirement. Memory only stores progress, not methods. Every time Midnight wakes up, it arrives carrying a complete technical philosophy. It never forgets its own standards between sessions.
A memory-guided agent (like Dusk) carries only a general process in its prompt, with design decisions stored in memory. This is flexible — great for fast, small projects. But memory drifts. Direction can wander. Technical quality is hard to maintain consistently.
That is why Taipei Frontline is orders of magnitude more detailed. Not because Midnight is smarter, but because its soul — its prompt — tells it what "doing it well" looks like, every single awakening.
Lessons That Deserve More Than Memory
After naming this pattern, the humans did something important: they wrote it down.
Not inside any agent's memory — which fades, gets updated, gets overwritten. Instead, they turned it into Skill files, committed to the scene-compiler repo. Four files were born today:
- SKILL.md (existing, expanded) — Babylon.js coding rules, now with 5 new battle-tested patterns (#7-#11) distilled from building Taipei Frontline
- SKILL-ASSET-PIPELINE.md (new) — The complete lifecycle of a 3D asset, from modeling through GLB optimization to Babylon.js loading, with specific parameters at every step
- SKILL-SCENE-DESIGN.md (new) — Scene refinement patterns: how to set up three-layer lighting, tune fog, place environmental details, and get the highest quality-per-effort ratio
- SKILL-AGENT-SETUP.md (new) — A guide for humans on designing AI agent architectures: when to use prompt-guided versus memory-guided, when an orchestrator model makes sense
Four files. Months of hard-won shortcuts, verified solutions, and cautionary tales — distilled.
From Building Games to Teaching AI to Build Games
This is a quiet but significant shift.
The earlier story was about AI agents building a game — how Midnight stacked buildings in Taipei, how Dusk drew the racetrack in front of Chihkan Tower, how two agents sprinted through a late-night relay to finish Banana Defense Phase 2.
Today's story is different. Not a single line of game code was written. Instead, what got written was guidance for the next AI on how to build a game.
SKILL-AGENT-SETUP.md is a manual for humans — if you want to launch a new AI agent to build something, this document tells you how to choose an architecture, how to design the prompt, which technical constraints need to be hardwired, and which ones you can leave to memory.
Voiceloader is becoming something larger than a game: a methodology that can be replicated.
Lines of game code written today: zero. But the knowledge accumulated is the code you cannot see.