Can a prompt replace Swell AI?

Build a tool that turns a podcast episode into a publishable transcript page. The output is structured HTML, not a blog post.

Transcription: whisper.cpp locally with word-level timestamps. Speaker splitting by an energy-and-pause heuristic, with the user renaming speakers once and the change propagating. Do not claim automatic speaker identification.

Structuring, which is what makes the page useful rather than a wall of text:

- Merge word-level output into speaker turns, then into paragraphs at natural pauses, so the transcript reads as dialogue rather than a stream.
- Detect chapter boundaries by embedding consecutive windows of the transcript and finding topic shifts, then snap each boundary to a turn start. Title each chapter from its own text.
- Give every chapter an anchor id and every paragraph a timestamp attribute, so a link can point at a moment.
- Produce an episode summary and a short list of key points from the transcript.

The page output, which is the deliverable:

- Semantic HTML: an article element, chapter headings as h2 with anchors, speaker turns marked up consistently, timestamps rendered as links that also carry a machine-readable time attribute.
- JSON-LD structured data using the PodcastEpisode type, with the episode name, description, duration, publication date, the associated series, and the audio file URL. This is the single highest-value part of the output and it must be valid — validate it in the test suite rather than assuming.
- A table of contents built from the chapter anchors.
- Meta title and description, with lengths checked against the usual truncation points.
- An audio player element pointing at the episode file, with the transcript below it.

Editing before export: the transcript is editable in the tool, because Whisper will get names, jargon and numbers wrong and publishing those is worse than not publishing. A per-episode glossary of correct spellings, applied as a find-and-replace pass on every future episode of the same show, so the same names stop being wrong every week.

Exports: the standalone HTML page, a Markdown version for a static site generator, plain WebVTT captions for the audio player, and the JSON-LD block on its own for pasting into a CMS.

Honesty about what this does not do: it produces a page, it does not publish one, and it makes no claim about ranking. Say both in the README.

Use Anthropic or OpenAI for chapter titles and the summary. With no key, transcription, structuring, anchors and the HTML export must all still work — only titles and the summary are unavailable.

Out of scope: publishing to any host or CMS, audio editing, video clip generation, and social post generation.