Bottleneck

Bottleneck

I run OTZ Group, a two-person AI implementation company in Miami, with my brother. On a normal day a bakery needs its delivery payouts reconciled, a customs broker needs an entry filed for a container at the port, and a dental office needs help with calls its front desk can't get to. Each of them wants me to handle it. Bottleneck is a game about that pressure. You play the founder at the point where one person can no longer cover every client by hand and has to build systems that handle the routine work.

You walk to a client's door and stay there until the job is finished. Some jobs are quick, some are large, and some grow while you are working on them. Leads appear during the quarter, and signing them adds revenue and more doors to cover. To keep up, you can hold Space at a door for three and a half seconds to build that client an AI agent. You can't move while you build, and every other client's timer keeps running. In my work, building a system always means something else waits, and I wanted the player to make that call under pressure.

Design decisions I would defend:

• Agents take the routine work. About one request in four still escalates to you, and escalations

pay more. In our engagements, automation moves my time to the harder problems clients value most.

• Scope creep. In an earlier version, reaching a door finished the job instantly. I changed that

because it misrepresented the work. Every job now takes time, and about 30 percent of jobs get longer while you are working on them.

• Growth is the player's choice. A player who never signs a lead keeps the quarter safe and ends

with about a third of the revenue.

• My brother can be called in once a quarter, and the clients speak English and Spanish,

because that is how our business runs.

The most important input to this game was built before the assignment started. Over the past year I built Co-CEO, a repository that holds the working knowledge of my company: client profiles, engagement records, meeting summaries, strategy documents, how we price and deliver, and how I write. Claude Code runs on top of it every day, and I built this game inside it. When I asked Claude to make the game reflect me, it worked from more than 700 files about my business and my decisions. The concept, the clients, the bilingual dialogue and the lesson at the center of the game all come from that record.

Claude went further than I asked in testing. It had a bot play simulated quarters and found that a perfect player could handle everything by hand and never lose a client, which would have undercut the point of the game. It tightened the pace until covering every client yourself costs you clients. Where I overruled it was the instant jobs. After playing, I knew they did not reflect how work actually goes.

I did not write the code, and I don't think that is where authorship sits anymore. These models write better code than I ever would, and that is the reason to use them. What I authored is the system that tells the model who I am, what my business does and how I make decisions, plus the judgment calls I made after playing the result. The game came out specific to my life because I spent the past year putting my company into a form a model can use.

Prompt that guided it

"One thing I'd change is to account for scope creep in the work. Some builds end up taking longer than others, and you can't instantly serve the client. You need to wait there longer. This is a more accurate representation of what work really is."

AI disclosure

Tool: Claude Code (Claude Opus 5.5), running inside Co-CEO, my company's knowledge base repository, in one session. I gave Claude the assignment and directed it to build a game that reflected me using the context in that repository. Claude proposed the concept, mechanics, art and on-screen text from that context and wrote all of the code. It ran about four rounds of its own testing and fixes, including 34 simulated playthroughs to set the difficulty. I played the game and asked for one change, real work time with scope creep, which Claude built and rebalanced with another 32 simulated quarters. Claude drafted this statement from my notes and the knowledge base, and I revised its framing and language. Nothing was abandoned.