A Scoundrel’s Tale is a story-driven roguelike card RPG. Five acts, twenty-five dungeons, four skill trees, an original score, full English and German, on iOS and Android. Act I is free; one purchase unlocks the rest. There is no studio behind it. There isn’t even a full-time developer behind it. I made it in my spare time, over roughly 2 months, and the honest headline is that I could not have made it alone. Most of it was built by AI, under my direction.
I want to be completely upfront about that, because I think the interesting story isn’t “look at my game.” The interesting story is what it means that a single person with a clear vision can now ship something like this at all, without the years, the team, or the budget.
Where it started
The seed is Scoundrel, a 2011 solitaire roguelike by Zach Gage and Kurt Bieg that you play with an ordinary 52-card deck. It’s a brutal, elegant fifteen-minute game, and it had a quiet resurgence recently: videos, print-and-play decks, little digital versions. I played it and couldn’t stop thinking about a bigger question: what if this loop had a life, a story around it?
That’s a vision, and a vision is cheap. Building it the traditional way (learning enough game dev, producing dozens of illustrated cards and backdrops, composing a score, writing and localizing a five-act story) would have cost years I don’t have. So this became an experiment: how far can one person get by directing AI end to end?
What the AI did, and what I did
The division of labor was consistent across every discipline: the AI produced, I directed, judged, and rejected. That second part is not a formality. It’s most of the work.
- Code: nearly all of it is written by Claude (Anthropic’s coding agent), in TypeScript on React Native/Expo. The game engine is pure, deterministic TypeScript with a seeded RNG and a test suite the agent maintains alongside every change. I set the architecture rules; the agent enforces them better than I would.
- Art: every backdrop, card illustration, and emblem is AI-generated from a hand-maintained art bible: one fixed style (layered cut-paper and dry gouache), one fixed palette, prompt by prompt. The keeper work is curation. For each asset I generated variants and threw most of them away until I was satisfied with the result.
- Music: the score is AI-generated (Suno) against a written score plan, one theme per dungeon, then analyzed, loop-cut, and transcoded with ordinary audio tooling. Takes that didn’t match the plan’s key, tempo, or mood got binned, same as the art. This was also the hardest part to get right, and still the one I’m least happy with (from a quality standpoint). I like the actual tracks, but Suno can’t really create a loop and I had to make the most of it with my limited audio skills.
- Story: the world, the characters, and what the story is about came from me and stayed under my control through written “bibles” the AI must follow. The prose itself was drafted by AI and went through repeated editorial passes, including sessions where the AI interviews me about what should happen next, and line-edit passes where I approve every changed line. I learned a lot from Arc 1 and will be able to improve this process for the next arcs, so you should see a jump in quality from there.
- German: I didn’t want machine-translation German, so the localization runs through a three-pass pipeline: meaning transfer, then a source-blind rewrite by a “translator” agent that never sees the English sentence structure, then a fidelity check. I approve every scene. I am a German native, which comes in handy here, but I hope I can eventually get to a point where the pipeline is robust enough to enable me to ship a game in a language I don’t speak. A key learning here: the translation can only ever be as good as the source material, so improving the quality of the English text is a prerequisite for improving the translated text.
If you total that up honestly: taste, direction, and final responsibility were mine; almost all execution was AI. I decided the game should exist and what it should feel like. The tools made the deciding matter.
What didn’t work
- Creating real loopable music with Suno.
- Letting AI translate the game without a multi-pass pipeline and human oversight. The result read like German on the surface, with English sentence structure and calques underneath.
- AI has no taste of its own. Left alone it converges on generic fantasy mush, in words, in pictures, in music. The bibles, palettes, and register rules exist because the default output is bland, and everything good in this game is the result of rejecting the first answer.
- Nothing ships itself. Store review, certificates, screenshots in ten sizes, privacy forms, an imprint page under German law. The unglamorous half of shipping is still very human.
Why I think this matters
There is a lot of understandable anxiety about AI-made things, and a lot of low-effort AI content justifying it. I can only offer this game as a counter-example: nobody typed “generate a game.” One person spent real care over a long time and used AI as the workforce their idea could never otherwise afford. The result is a game with a specific, personal thing to say. Wren’s story is about grief, and it goes places I haven’t seen a card game go.
If you have an idea like that in your head (a game, a tool, a book), the gap between “I can imagine it” and “it exists and strangers can buy it” has never been this small. That, more than my game, is the thing I’d want you to take from this post.
A Scoundrel’s Tale is out now on the App Store and Google Play. Act I is free. If you play it, I’d genuinely love to hear what you think: hello@scoundreltale.com. Questions about the process are welcome too; I’ll answer anything.