I had a peculiar combination of events. Anthropic gave me 20$, then 100$ of Extra Usage credits, same as any other Pro subscriber. It’s promotional, and it’ll expire somewhat soon. I wanted to use it up.

I read Fable is phenomenal at very long horizon tasks, and it drains Extra Usage on my Pro plan. I also had a project I’d abandoned a while back - it was started when gpt-oss-120b was just released, and tool calling was still somewhat of a novel thing on a model I could just use for free via various platforms like Groq. I’d wondered how a story could be driven with a “multi-agent” (read: 2) architecture of a narrator and a companion character. They’d have mostly full control of the story, and the player could say anything, and have their own story. Because, well, LLMs.

Unfortunately for the project, I got bored… very very early on, and I ended up leaving it half broken, untested, and with just the start of a story drafted. It was the perfect project to point Claude Code at, switch the model to Fable, and quite literally ask it to drive the project to its logical conclusion.

Here’s what I asked:

Hello! I have a slightly ambitious, and possibly a bit of a stupid request for you. This is an abandoned project, made back when gpt-oss-120b was brand new and I was just exploring how tool calling (now mostly advertised as agentic capabilities) might be used for stories.

I want you to create a branch, then use dynamic workflows to offload to Opus and Sonnet, to drive this project to the logical conclusion, including everything from updating any old packages or assumptions, and finishing the story entirely. You may change design parameters too, to better fit the reality and better capabilities of modern models. Do you think you can do that?

It then proceeded to run for 29m59s, using up 3.71$ of Extra Usage credits and about half my session quota, albeit with a 1.5x promo on Claude Code’s weekly usage running. After I poked at it a few more times to assess the story quality, expand the LLM context windows, and playtest the thing, the project seems ready to play again - I’ll probably play through it when I get the time to.