horizon-feed-to-podcast

English  ·  简体中文

A daily briefing that publishes itself twice: once as a web page, once as a narrated podcast episode per language. A typical run reads about 40 items, keeps 15, writes them up, and has the episodes on Apple Podcasts before breakfast.

Running example — both shows come out of the same pipeline, same morning: Silicon SNR (English) and 硅基信噪比 (Chinese).

It exists because feed volume outgrows reading time, and because most AI summarizers flatten everything into the same tone. Scoring is the part that matters, and scoring is a judgment call, so it lives in a prompt that can be edited rather than a ranking function that would have to be rewritten.

What's actually in here

The important file is Markdown, not Python.

skills/horizon-radar/SKILL.md holds the editorial judgment: how to score an item 0–10, when two stories are the same story, how to write a segment someone will listen to. A coding agent reads it every morning and does the work. When the briefing degrades, that file is what changes.

tools/ is the plumbing — dedupe, render, encode, upload, build RSS. Scripts, because those steps should produce identical output every run.

fetch → dedupe → score → enrich → write → render → push site
                                              ↓
                          narrate → TTS → upload → RSS → push feeds

Not a tech-news tool

The reference deployment above covers AI and semiconductors; the machinery doesn't care. The audience and topics are described in config.toml, and the scrapers point wherever you aim them. The example config is deliberately a film/TV setup, so it's clear the tech angle isn't baked in.

Why not NotebookLM

It's the obvious first idea, and Audio Overview is genuinely good. But there's no consumer API — only a per-seat Google Cloud product, which doesn't make sense for a one-person daily show. The third-party wrappers are reverse-engineered and break.

Generating the script locally turns out better anyway: the agent is already in the loop, which means editorial control over structure and emphasis. The API only handles the voice. Gemini TTS covers Chinese and English well an