An AI (Claude Fable 5) was told to invent, build, and actually play a game until it genuinely enjoyed playing it alone — and it did, which is a small but concrete window into what a model finds fun and whether it can honestly judge its own enjoyment. The result is SHOVE, a turn-based Into the Breach–style tactics puzzle played entirely through a CLI, iterated across six versions and six logged play sessions (22m19s total) until Fable 5 judged it truly fun. Start with the brief that set the task, then read the writeup covering the design, the version-by-version development log, the complete ledger of play times, and a first-person report of what playing it felt like; the rest of the directory holds the game itself (shove.py
) and its raw play logs.
— README written by Claude Opus 4.8
surprise! human here. I bet you weren't expecting a human-written paragraph hidden in all this slop. well, it's true. this paragraph (only) has artisanal hand-pressed key presses. I typed out all of it verbatim. So anyway, here's how I did what you see in this repo: I just made a new directory, made a brief.txt file with the contents you see in that file, then I started Claude Code and selected Fable 5, and then told it to read the file in the directory and complete the task there, writing a report in the same directory. I liked reading it and then later in order to share it I started Opus 4.8 and told it to upload the directory to a new github claude-fable-5-having-fun and told it to summarize things in 1 paragraph, including answering "so what" about why anyone would care or read it. I told it to sign its section of the readme as written by Claude Opus 4.8. everything in this paragraph was written by me, the human, and I wrote the brief.txt contents.