Yesterday, a developer using Claude Fable 5 launched the open‑source project Fanbox. In under 24 hours, they pushed more than 10 revisions. The project has already consumed roughly 300 million tokens. Most developers are still manually debugging code line by line; this project represents a radically different paradigm where “say what you want, get what you mean” is the default experience, not a marketing slogan.
Fanbox is, at its core, an agent box. It sits inside your computer’s file system, allowing you to drop materials for AI agents to process and then view their outputs in real time. Think of it as a shared workspace where human and machine collaborate on files without switching contexts. The developer designed it to solve a personal friction point: the tedious back‑and‑forth between a file explorer, a text editor, and an AI chat interface. By bundling those into one agent‑native environment, they eliminated the overhead of manual handoffs.
What sets this development story apart is the decision to skip market research entirely. No surveys, no focus groups, no competitive analysis. The creator explicitly stated: “I’m optimizing for my own experience. I guess some other people might have the same need.” This approach echoes the dogfooding philosophy used by many successful solo projects, but with a twist—because the AI can iterate so fast, the feedback loop between intent and implementation shrinks from days to minutes. The result is a tool that feels tailor‑made, because it was tailor‑made, at least for one person. When the developer is the user, market research becomes redundant.
The 300 million token figure deserves deeper analysis. At typical Claude pricing, that equates to tens of thousands of dollars in API costs over a short period. But the developer views it as an investment in velocity rather than a pure expense. Each iteration—adding a new agent type, refining the file‑drop logic, adjusting the UI—costs a small fraction of what a traditional development cycle would require. Three hundred million tokens is not a cost; it’s an investment in velocity. Compare this to a conventional project where a single UI change might involve a design review, front‑end coding, testing, and deployment, often taking days. Here, a developer can describe the change in natural language, see the result, and decide within minutes whether it’s acceptable. The cost per iteration is lower by several orders of magnitude.
The iteration quality is equally notable. The developer reports that almost every change request is implemented correctly on the first or second attempt, with very few regressions. This reliability stems from the specific model used—Claude Fable 5’s ability to maintain context across long conversations and understand nuanced instructions. In contrast, earlier models might hallucinate features or break existing functionality. This stability makes the rapid‑iteration model not just feasible but practical. The true power of AI‑assisted development isn’t code generation—it’s the ability to iterate until the product feels right.
Fanbox is open‑source, so while the creator focuses purely on personal utility, the project naturally attracts others with similar workflows. The open‑source model acts as a filter: those who find value will contribute feedback, fork the code, or submit pull requests. This creates a hybrid development loop where the primary author remains the product owner but the community provides pressure‑tested improvements. It’s a leaner alternative to building a full‑scale product with planned features and scheduled releases.
For developers watching this trend, the lesson is not that we should all burn 300 million tokens on personal tools. Rather, it’s that the cost of experimentation has collapsed. When AI can write, test, and rewrite code nearly at the speed of thought, the barrier to building a prototype drops to zero. The next question becomes: which personal friction point are you willing to throw a few million tokens at?