From any document
to a produced voice & video story.

Everything on this site is the artifact of a deliberate under-specification experiment: an empty directory, one prompt — "create a web API that takes a URL or PDF and converts it into a podcast" — and fewer than 500 human words of direction in total, while Claude Code (Fable 5) made the engineering decisions. The author works professionally on complex agentic harness development across different environments — here, deliberately, stock Claude Code with no extra add-ons was the harness, running Fable 5, the most capable model publicly available — the frontier, not a budget pick — for the most honest impression of what it does on its own. Roughly 3,400 lines of working code, 17 endpoints and this studio emerged in a single evening; the code review that redirected its architecture was four words long. What you are browsing demonstrates a way of working — the code is experimental, not production-ready: not security-hardened, not load-tested, not something to deploy as-is.

The finding cuts the other way too. Even at the frontier — as capable as publicly available AI currently gets — the model is no holy grail. Not once did it propose a direction or say "this is enough"; it wrote thousands of lines and zero tests; twice it reported success on changes that silently did nothing; and it missed flaws any user sees in seconds. Everything it lacked — taste, dissatisfaction, verification, standards — had to come from the human side of the prompt. That is the experiment's real result, and it is a direct challenge to the executive dream of automating everything with AI: capability can be rented; judgment cannot. The human contribution is not disappearing — it is concentrating. Read the experiment write-up on LinkedIn →