Podcast Highlights Deck

(Internal skill id: podcast-highlights-deck)

What it creates

Create a premium editorial long-scroll highlight deck with a sticky TOC rail, multilingual toggle, and original audio clips.

Workflow

Inputs

• podcast_url: episode page URL (Apple/Spotify/RSS/YouTube/direct MP3)
• languages: list of language codes, e.g. en, ja, zh

Output

• A static website (Vite build) with:
• editorial hero (no full-bleed podcast artwork)
• sticky left rail: metadata + language toggle + table of contents
• 8–12 highlight sections
• per-highlight audio clip playback (original audio)
• global language switching (no mixed-language UI)

Workflow (execute in order)

1) Acquire audio (source-of-truth)

Prefer a direct audio URL (RSS ). Recommended approach:

1. Use search_web to find the show’s RSS feed (queries like: "" RSS feed, or the Apple show id + RSS).
2. Use get_web_page_contents to fetch the RSS XML.
3. Parse RSS to locate the exact episode and extract:
• title
• publish date
• duration (if present)
• cover image
• enclosure mp3 URL

If RSS is unavailable:

• If YouTube exists, use yt-dlp to download audio.
• If a platform blocks direct audio access, ask the user for the RSS link or direct mp3.

Download audio to a working folder (example):

• podcast_work/episode.mp3

2) Transcribe with timestamps

Primary:

• Run anygen-speech-to-text episode.mp3 -o transcript -f json,md,srt.

Fallback (if the tool fails):

• Split audio into chunks with ffmpeg (10 min chunks)
• Use OpenAI Whisper (whisper-1) with response_format="verbose_json"
• Merge segments by adding time offsets

You need a machine-readable file like:

• transcript/episode_verbose.json containing segments with start, end, text

3) Curate 8–12 highlights (do NOT dump transcript)

Selection philosophy:

• Prefer fewer, stronger highlights.
• Only use quotes that exist in the transcript.

For each highlight, produce:

• id (h1..h12)
• start + end timestamps in seconds (from transcript)
• title (translate later)
• quote (English, exact or lightly cleaned)
• context (1 sentence)
• takeaway (editorial interpretation)

4) Translate + global UI copy

For every supported language:

• Translate titles, context, takeaway, and quote (transcreation; keep meaning + tone).

Important behavior:

• In non-English modes, show translated quote as primary.
• Preserve a connection to English:
• show “Original (English)” as a secondary expandable panel.

Also translate all UI strings:

• hero framing
• sidebar labels
• buttons (“Play clip”, “Back to top”, etc.)
• closing section labels

5) Clip original audio per highlight

Use the bundled script:

• python scripts/clip_audio.py --audio episode.mp3 --highlights highlights.json --out-dir site_assets

Conventions:

• add ~2s padding before/after for natural listening
• output:
• site_assets/audio/h1.mp3 …

6) Build the site with the bundled editorial template

Use website_init to create a new site project.

Then copy assets into the project:

• src/assets/highlights.json
• src/assets/cover.jpg
• src/assets/audio/*.mp3

Then replace template files from this skill:

• assets/template/Home.tsx → src/pages/Home.tsx
• assets/template/index.css → src/index.css
• assets/template/index.html → project index.html

Notes:

• The template expects highlights.json schema similar to assets/template/highlights.schema.example.json.
• Ensure document.documentElement.dataset.lang is set from the language toggle.

7) Bundle and deliver

Run website_bundle and deliver the generated dist/index.html.

Template assets in this skill

• assets/template/Home.tsx: editorial layout + global language switching + expandable English original
• assets/template/index.css: Swiss‑brutalist paper/ink theme + language font stacks
• assets/template/index.html: Google Fonts includes Instrument Serif, Manrope, IBM Plex Mono, and Noto JP/SC
• assets/template/highlights.schema.example.json: reference structure