Books

A podcast that writes itself, documented by the person who has to keep it running. Someone sends a question; about twenty-seven minutes of finished audio comes back. Nobody hears an episode before it goes out. There is no review step and no approval queue.

This is the production documentation for that system as it ran on 10 August 2026 — the ingestion path, the five-node generation graph, model routing, voice cloning on GPU workers, the memory layer, audio assembly and publication. Then it prints the prompts themselves, in full, annotated line by line.

Most writing about agentic content pipelines describes the happy path. This one describes day two hundred: an advertisement mechanism running on its emergency fallback for months, a version column that identifies almost nothing, an interface built to replace an older one and never connected to anything. None of them stopped the show going out. That is the uncomfortable part.

Every episode of a podcast that publishes every day, in one place you can hold. At 4,620 episodes — about 2,107 hours, or 88 days of continuous audio — the show has a problem no amount of search solves: you cannot want an episode you do not know exists.

This is the index. Every episode, in publication order, grouped by month, each with its number, its permanent six-digit code, its runtime, its publication date, its category, and one line telling you what it is actually about.

It is not a highlights collection and it is not curated. There is no "best of" here and no editorial. Completeness is the entire proposition: if an episode exists, it is in this book, including the ones nobody should listen to.

Find something, then type six digits. Every entry carries a short code — a permanent identifier that is never reassigned, even if episodes are withdrawn and the numbering shifts around it. Put it after the site address and it opens the episode directly: myweirdprompts.com/708787. The six digits are the whole interface.