Carbs of the Huge Manatee, or CHM, began as Clint Zehner's series of Unauthorized Expansions for Cards Against Humanity. He created nine expansions of roughly a hundred cards each, sold them through Amazon, and came close to placing an expensive bulk order immediately before receiving a cease-and-desist notice. The project later became a useful measure of how far AI-assisted software development had progressed: an online game that had defeated his early attempts became much easier to build when he returned with more capable tools.1

Writing and selling the cards

The original project involved making additional material for a game people already knew how to play. Clint wrote and produced nine expansions, each with approximately a hundred cards. His personal favorite was Online Dating 1, one of the sets preserved in the project's card archive.1 2

Amazon provided a way to sell the decks beyond the people who could play them with him. That turned writing cards into a small product business, with the practical work of producing enough decks to keep selling them.1

The largest near-miss came when Clint ordered three lots of a thousand new decks for Amazon, at a cost of almost $10,000. The following day, he received a cease-and-desist notice from Cards Against Humanity. According to Clint, the objection concerned the resemblance of some older decks to the official game's design.1

Fortunately, the bulk order had not yet been processed. He was able to cancel it. The timing left him with the end of that sales chapter rather than three thousand newly produced decks arriving after the notice.1

The first online attempts

The cards also suggested another project: an online version that friends could play together. Clint had the game and a clear idea of what it needed to do. With AI beginning to help write software, it seemed like something he might be able to build.1

The early attempts were frustrating. He worked through multiple versions with an early ChatGPT model, repeatedly having to remind it of the variable names already in use. The difficulty was keeping the growing project consistent. A useful-looking answer in one part of the conversation did not necessarily fit the code and decisions established elsewhere. Eventually, he gave up.1

The unfinished work remained. Its archived plans described the players, rooms, decks, judging, scoring, and chat in detail. A later review found those design records useful even though the old implementation could not be carried forward as a working game.3

Returning to the project

When Clint tried again with a more advanced AI model, he was stunned by the difference. The same overall project was now relatively straightforward to develop. That did not remove the need to understand the game or make design decisions; it made it possible to connect those decisions into a functioning application without constantly repairing the conversation's memory of its own work.1

The rebuild retained the useful design work and replaced the unfinished implementation. By early September 2026, the project records described a complete game loop that had been exercised through finished games, including human and Rando winners.3 4

Some of the challenges were specific to translating a card table into a browser. Players needed to find a room and return to it. The person judging a round needed to see the answers without being shown who played them. Answers using two or three cards had to keep the player's intended order. A game also needed to handle someone pausing or leaving, award the final point, and let the table start again afterward.4

The finished-game work mattered because dealing cards and accepting an answer were only parts of the experience. The application had to keep the same game coherent from the lobby through the last round. The documented rebuild covers that full sequence.4

The online version

CHM Online is a private browser game with a lobby, searchable room names, invitations, deck selection, and host controls. Players answer prompt cards from their hands, while the rotating Card Czar reveals anonymous submissions and chooses the winning response.4 5

Submissions stay concealed while they are being collected and are shuffled for judging. The game supports ordered multi-card answers, room chat, game history, pause and resume, and a new game at the same table after the previous one finishes.4

Rando Cardrissian adds an automatic response from the deck. Rando can score points and win, but is distinct from the planned AI opponents. The documented current game requires two active human players.4 5

Future features

Clint plans to add AI players with different personalities, allowing a person to play solo against AI opponents. That would add another role for AI in the project's history: first an unreliable development helper, then a much more useful collaborator, and eventually a participant at the card table.1

The solo-play and personality features remain future additions. They are separate from Rando's existing automatic card submission.1