Machinations: the game that uses dice, not the model, to force AI to narrate failures

By Hacker News (Show HN) · July 21, 2026.

A user identified as jklewis has presented on Show HN "Machinations", a multiplayer strategy game set in a court of intrigue —the author himself compares it to Game of Thrones: conspiracies, betrayals and murders to seize a throne— in which a language model (LLM, the type of AI behind ChatGPT or Claude) acts as referee or "game master", the role that in tabletop role-playing games decides what happens when a player tries something.

According to the author, the starting point was not the game but a technical problem. He comes from "diceless" role-playing games, where what he valued most was the human referee improvising resolutions on the fly. When trying to transfer that role to an LLM, he ran into a known limitation of these models: they tend to "regress to the mean" and to over-please the user, so that, left to their own devices, almost no one ever fails in the story. Without failure there is no conflict, and without conflict there is no game.

The solution he describes is purely mechanical: a 3d10 roll system (three ten-sided dice) that generates a bell curve of possible outcomes, from "disastrous" to "legendary". The LLM does not decide whether the player succeeds: it can only narrate the outcome the dice have already determined. On top of that, the author added a classifier that analyzes the narrative situation and decides which character has an advantage or disadvantage, translating it into modifiers on the roll. Finally, a backend process (the part of the system the player does not see) dissects the narrated text to extract what actually changes in the game state, so that fiction and rules do not become decoupled. The published game mode —conquest of a kingdom— is the first of several he plans to add, since each mode is, by design, just a different story to be resolved with the same engine.

It is a modest launch: at the time of writing, the post had a single vote on Hacker News and was less than an hour old, and the author did not even want to name the system to avoid it sounding like self-promotion. There is no company here, no funding round, no user figure to analyze. What there is, is a small but sharp case study of a big problem.

The tendency of LLMs to please —to avoid conflict, the "no", genuine failure— is not a quirk of this game: it is a well-documented consequence of reinforcement learning from human feedback (RLHF), the process that tunes these models so their responses appeal to whoever evaluates them. The same bias that makes a coding assistant say "great idea!" to almost any proposal, or an AI tutor rarely fail anyone, is what here prevented a character from dying in combat when the story called for it.

This developer's solution is interesting precisely because it does not try to fix the model with more prompt: it takes away the LLM's power to decide and gives it to a number generator external to it, leaving the AI to merely narrate and interpret consequences. It is the same pattern beginning to be seen in more serious agentic systems —separating the part that decides (deterministic, auditable) from the part that expresses itself in natural language (the LLM)— because delegating weighty judgments to a model that by design wants to look good is, simply, delegating them badly.

Our reading is that cases like this, however small, say something real about the state of the art: the demonstrated ability of current LLMs to sustain genuine narrative conflict, without external help, remains limited, however much "creative AI" marketing suggests otherwise. The practical path, today, is engineering around the model —dice, classifiers, hard rules— and not the model on its own. It is also a useful reminder that the frontier of what can be built with AI no longer requires a large team: a single developer, after years devoted to a system of his own, can today launch something like this on his own. That accessibility, more than the game itself, is the part that connects with an underlying trend worth following: the cheaper and more accessible it becomes to build with these models, the wider the range of who can attempt it.