3D Tetris — One Shot, Full Game

I wanted to test out DeepSeek-V4-Flash Q2 to see what its capable with a one shot prompt, using a RTX 3090 — 24GB VRAM, 96GB DDR5 RAM. It runs at 13.5 tokens/sec which is quite slow and took around 30 minutes to create the code. I gave it a single prompt describing a 3D Tetris game. What came back was a complete, playable game with SRS rotation, 7-bag randomizer, ghost piece, hold, scoring, and bloom post-processing — all coded in one shot, running locally on one RTX 3090.

The Prompt

Build a playable Tetris game rendered in 3D using Three.js. The game should look visually appealing with 3D blocks, proper lighting, and standard Tetris gameplay mechanics.

That's it. No architecture doc, no tech stack specified beyond Three.js, no wireframes. One paragraph describing what I wanted to play.


The Simulation

Arrow keys or WASD to move, space to hard drop, Z to rotate counter-clockwise. Ghost piece shows where you'll land. Hold C or Shift to swap a piece aside.


What It Does

The game implements 16 distinct features:


The Numbers

978
Lines of code
7
Tetrominoes
4
Pipeline agents
1
Coding shot

Built with Three.js r184 loaded from CDN. Single HTML file. No build step, no dependencies to install. Runs in any modern browser.


How It Was Built

This wasn't a single model call. The PI Agent Harness ran a 4-agent pipeline — each agent handling one job, coordinating through structured prompts and a shared task board. All running locally on a single RTX 3090.

The Pipeline

Researcher Reviewer DeepSeek-V4-Flash (Coder) Reviewer

What Each Agent Did

  1. Researcher — Researched Three.js capabilities, SRS mechanics, wall kick tables, collision patterns, lighting, post-processing (32KB research report with source citations)
  2. Reviewer — Validated the research, confirmed the approach was sound before coding began
  3. DeepSeek-V4-Flash (Coder) — Wrote the complete 978-line implementation in ONE SHOT: game logic, rendering, input, effects, UI. The coding was done in a single call by DeepSeek-V4-Flash Q2 from Unsloth, running locally on one RTX 3090.
  4. Reviewer — Final audit: 12/12 criteria verified, verdict: complete

Why It Matters

This is what a local AI agent team can ship. No API keys, no cloud credits, no vendor lock-in. One prompt describing a game I wanted to play — and a pipeline of specialised agents that researched, designed, built, and reviewed the whole thing on a single consumer GPU.

The model that drove the coding was DeepSeek-V4-Flash — a 284B-parameter Mixture-of-Experts model with only 13B activated parameters per token, running quantised (Q2 GGUF) locally. It has a 1-million-token context window and an MIT licence.

The model is available on HuggingFace and the GGUF variants at HuggingFace GGUF.

The harness that made this possible is documented in the PI Agent Harness bundle — ebook, video series, and all agent definitions.