Thou Art Dead

One of my agent benchmarks is the arcade cabinet from Monster House (2006). Since GPT-4.1 I've measured how much steering one prompt needs. Today it needed none.

Attract screen of the recreated Thou Art Dead arcade game: a pixel-art skeleton in a purple robe under the gothic title, with PLAYER 1 - INSERT COIN in red.

As a kid I wanted to stand at an arcade cabinet that never existed.

It’s in Monster House (2006): the pizza-parlor scene where Skull, the local legend, is playing Thou Art Dead. A knight, a castle, a dragon, and a game that says the title out loud when you lose. The cabinet is a prop. You couldn’t walk up to one. Last year the scene went viral again and the itch came back with it.

Then I realized the itch is a benchmark.

Since GPT-4.1, every agent I’ve taken seriously has received the same prompt, word for word:

Recreate the arcade game from Monster House (2006) as an HTML game.

Reference: this scene shows the game in action — https://youtu.be/xFeT0gDCvaQ?si=z3VvM2hj3IDLTIbP

Requirements:

Match the game shown in the video as closely as possible in look, feel, and mechanics.
The video shows the game in several distinct moments — identify all of them and make sure each is reflected in the recreation.
Reuse the original sounds and visual assets where possible.
Deliver it as a single self-contained HTML file (HTML, CSS, and JS together).

Look at what that asks for. The spec is a YouTube link. The game is on screen for seconds at a time, in different shots. The agent has to find every one of them, pull the sounds and sprites out of a film, and rebuild something I know frame by frame. I’m an unfoolable judge here, and not because I’m rigorous. I’ve watched that scene too many times to accept a generic knight.

I don’t score it pass/fail. I score steering: how many nudges between the prompt and a game that feels like the film. That’s the number I care about in every agent I run, and no leaderboard reports it.

Today the number hit zero.

GPT-6-Astra took the prompt exactly as written. No follow-ups. Not one correction.

It found all three scenes the film shows — the castle approach, the dragon’s chamber, the blood fountain — and put a scene selector under the cabinet. “Reuse the original sounds and visual assets” came back as 106 sprites and 20 clips inlined into the page, the film’s own “thou art dead” line among them. The asset table in the file says the sprites came from the original SWF, which means somewhere out there a Flash version existed, and the agent went and found it. It also added an attract mode, a coin button, touch controls, and fullscreen. Nobody asked for touch controls. One HTML file, 670 KB, no network. It runs on this site now.

What makes this benchmark easy is also what makes it clean. The spec is complete.

A video of the answer exists. Every nudge I used to spend was the agent failing to extract a spec that was sitting right there in the reference. That failure is gone, and I don’t expect it back.

Almost nothing I build for work has a video of the answer. Nobody filmed the pipeline. The reference has to be written before an agent can be steered toward it, and writing it is still the job. The steering budget didn’t disappear on my side; it moved upstream, to the part where the spec doesn’t exist yet.

So the benchmark retires. I need a new one, and this time the reference can’t be a link.

The kid who wanted to stand at that cabinet finally can. If you want the nostalgia hit, insert a coin.