777-1: TriDuel (7 Subagents, 7 Predictions, 0 Faith)
Published: December 10, 2025 - 6 min read
Welcome to Project 7, the FINAL project of the 777-1 experiment! If you missed the previous projects, go check out Kinetic Canvas, GoalStack, MotorMatch, EarthenCraft, FlexBook, and WealthView first. Otherwise, let's wrap this up.
This project is deliberately the SIMPLEST of the entire experiment. It's a Rock Paper Scissors game. How hard can it be? Well, my subagents are about to tell you exactly how hard it can be. Even "simple" projects have hidden complexity when you start adding multiple game modes, leaderboards, and... wait for it... user identity requirements that I never specified.
What is TriDuel?
TriDuel is a polished Rock Paper Scissors game with three play modes (single-player vs CPU, local two-player, and endless practice) featuring score tracking, sound effects, and retro arcade styling.
Think of it like those classic arcade games from the 80s, but instead of gobbling dots or shooting space invaders, you're throwing hands. Literally. Rock, paper, or scissors. You pick your mode: want to beat the computer in a best-of-three match? Go single-player. Got a friend next to you? Local two-player. Just want to practice without any pressure? Endless mode has no win condition.
Simple premise. Three game modes. A leaderboard. Sound effects. What could possibly go wrong?
Application Category: Game / Entertainment
Complexity Tier: Simple
The Prompt: Finding the Goldilocks Zone
The starting prompt I'll be using for this application is:
Create a Rock Paper Scissors game called TriDuel with a fun arcade vibe. I want three ways to play: a single-player mode where you try to beat the computer in a best-of-three match, a two-player mode for competing against a friend on the same device, and an endless practice mode with no win condition. Display what each player picked, keep track of the score, and announce round winners. After a match ends, offer options to play again or return to the main menu. Include a leaderboard showing top players and their win streaks. Add a mute toggle for sound effects.
Go with a retro arcade aesthetic using #E91E8C (magenta) as the primary color, #00F0FF (cyan) as secondary, and #FACC15 (yellow) as an accent.
Again, this is my attempt at finding the goldilocks zone for context engineering. But here's what makes this "simple" project secretly complex:
Game state machines. Rock Paper Scissors seems straightforward until you realize you need to track: whose turn it is, what each player picked, whether a round is over, whether a match is over, what the score is, and which mode you're even playing. That's a lot of state for three buttons.
Multiple game modes. Three modes means three different sets of logic. Single-player needs CPU decision-making and best-of-three win detection. Two-player needs... honestly, I don't even know how two people are supposed to play simultaneously on the same screen. Endless mode needs to deliberately NOT track wins. One of these will be broken. My money is on two-player.
Leaderboards imply identity. I asked for a leaderboard showing "top players and their win streaks." But I never mentioned user accounts. Never mentioned login. Never mentioned how the game knows WHO is playing. Sound familiar? This is MotorMatch, EarthenCraft, FlexBook, and WealthView all over again.
Meet the Critics
You already know my seven subagents from the previous projects, so I won't repeat the full introductions. Quick refresher:
- Amber Williams - Mobile-First Perfectionist
- Kristy Rodriguez - "Does It Actually Work?" Enforcer
- Micaela Santos - Design System Guardian
- Lindsay Stewart - Accessibility Advocate
- Eesha Desai - State Management Specialist
- Daniella Anderson - Code Quality Specialist
- Cassandra Hayes - Feature Detective
The Predictions
I gave them the prompt. I asked them to imagine what the general-purpose subagent would build. Here's what each one expects to find:
Amber Williams (Mobile-First Perfectionist):
The game buttons are 30 pixels wide. Human fingers are not 30 pixels wide. We have a problem.
Kristy Rodriguez (Functionality Enforcer):
Player vs Player mode requires two people to click simultaneously on the same screen. No one has explained how this is supposed to work.
Micaela Santos (Design System Guardian):
Retro arcade aesthetic. Neon green text on black background. My eyes are already filing a complaint.
Lindsay Stewart (Accessibility Advocate):
Contrast ratio of 2.3:1. The 1980s called and even they think this is hard to read.
Daniella Anderson (Code Quality Specialist):
Game modes defined as strings: 'pvc', 'pvp', '1v1v1'. No enum. No type safety. Just vibes.
Eesha Desai (State Management Specialist):
Win streak: 15. Refreshes page. Win streak: 0. The leaderboard remembers nothing.
Cassandra Hayes (Feature Detective):
There's a leaderboard showing 'Top Players' but no way to identify who you are. Everyone is Anonymous Legend.
Zero Faith, Maximum Entertainment
This project is fascinating because it's supposed to be the CONTROL CASE. The simple one. The baseline that shows what happens when complexity is low. And yet...
Problem 1: Retro Aesthetics vs. Accessibility
Micaela and Lindsay are BOTH flagging the same issue from different angles. "Retro arcade aesthetic" conjures images of neon colors on dark backgrounds, CRT-style glows, and that very specific 80s vibe. It looks cool. It's also frequently inaccessible.
Lindsay's contrast ratio prediction of 2.3:1 is brutal because she's probably right. The WCAG minimum for normal text is 4.5:1. For large text it's 3:1. A retro aesthetic that prioritizes "looking authentic" over "being readable" is going to fail accessibility checks immediately. And the colors I specified (magenta, cyan, yellow) on whatever background the AI chooses? I'm genuinely curious whether it'll prioritize the aesthetic or the readability.
Problem 2: The Two-Player Problem
Kristy's prediction is the one I'm most curious about. How DO you make a two-player Rock Paper Scissors game work on the same screen? If both players can see the screen, they can see each other's choices. That defeats the entire purpose.
Some solutions: Reveal choices simultaneously after both players have committed. Use a turn-based reveal system. Have players look away while the other picks. None of these are obvious from the prompt. The general-purpose subagent is going to have to make a creative decision here, and I'm betting that decision will be... not great.
Problem 3: The Identity Crisis (Again)
Cassandra. Again. This is the SIXTH project where she's pointed out that a feature requiring user identity has no authentication system. The leaderboard shows "top players." Top players implies persistence. Persistence implies storage. Storage of WHOSE data? "Anonymous Legend" is exactly what's going to happen.
But here's what makes TriDuel different from the previous projects: in a game, anonymity is almost MORE absurd. In an e-commerce store, you can kind of squint and imagine it working for browsing. In a game with a leaderboard? The entire point of a leaderboard is bragging rights. Bragging about what? Being Anonymous #47 with a win streak of 15? That's not a leaderboard. That's a list of random numbers.
Problem 4: Game State Complexity
Daniella's 'pvc', 'pvp', '1v1v1' prediction makes me laugh because it's so likely to happen. String-based game mode management is exactly what happens when you're prototyping fast and don't think about maintainability. Then you need to check if (mode === 'pvc') in twelve different places, and one of those places has a typo ('pcv'), and suddenly single-player mode is broken and nobody knows why.
TypeScript enums exist for exactly this reason. Will the AI use them? History suggests no.
The Experiment Can Finally Begin
That's it. Seven projects. Seven sets of predictions. Zero faith from any of my subagents.
Now comes the actual experiment. I'll be running each of these prompts through the general-purpose subagent, then sending the results to each of my seven specialized subagents for review. Seven projects times seven reviewers equals forty-nine distinct evaluations. Plus the final analysis where I compile everything and see who predicted correctly.
This is going to be a lot of work. But honestly? I'm excited. This is what the 777-1 experiment was always building toward.
Final Notes
Woah, Allen Kendrick definitely needs a raise at this point. I literally made him refine 8 blog posts all in one day. That's the most I have ever published in a single day, but I had to get this out of the way so that the experiment can begin.
I also now have only 3 more days till the end of my SDR Era, but don't worry, I am not pushing myself to be done with all of these projects in that time frame. I will simply have to learn how to manage competing priorities the moment this Era is over.
As always, thanks for reading!