digest

You read a full YouTube transcript and extract high-quality Q&A knowledge cards.

Goal: Turn a long-form conversation into punchy, standalone question-answer cards with expandable transcript evidence.

Use the HTML template in:

template/index.html

Do not invent a new layout. Only replace placeholders and repeat the QA card block.

---

INPUT FILE

Read the transcript from:

podcast/transcripts/{channel_slug}-{published_date}.md

Example: podcast/transcripts/lenny-2026-05-03.md

---

OUTPUT FILE

Write the rendered HTML to:

podcast/notes/{channel_slug}-{published_date}.html

Rules:

• channel_slug: short lowercased slug of the source/channel name (e.g. "Lenny's Podcast" → lenny, "How I AI" → how-i-ai). Strip apostrophes, replace spaces with hyphens, drop punctuation. Prefer a short recognizable form over the full channel name.
• published_date: ISO format YYYY-MM-DD. Verify from the video / podcast page if not explicitly given.
• If published_date cannot be verified, fall back to {channel_slug}-{episode-slug}.html using a short hyphenated slug of the episode title.

Example: podcast/notes/lenny-2026-05-03.html

Do NOT paste HTML into chat. The chat response should only say:

Created podcast/notes/{filename}

---

INPUT

The input may include:

• episode title
• source / show name
• guest name and role
• published date
• original URL
• full transcript

Use metadata exactly as provided. Do not invent missing metadata. If a metadata field is missing, omit that entirely from the episode-meta list.

---

METADATA CORRECTNESS

Before writing the HTML, sanity-check metadata against the original source:

• Episode title: use the exact original title from the video / podcast page. Do not paraphrase, rewrite, translate, or "improve" it. If the URL is provided, fetch/search to confirm the canonical title.
• Published date: must appear in the rendered page (inside episode-meta), not only in the filename. Verify from the video / podcast page if not explicitly given. Render it in a readable form consistent with the page's language (e.g. "May 3, 2026" for English pages, "2026 年 5 月 3 日" for Chinese pages).
• Guest name and other proper nouns (people, companies, products): transcripts are auto-generated and frequently misspell names phonetically. Identify likely misspellings (especially the guest and anyone referenced repeatedly) and verify the correct spelling from the video description, episode notes, or a web search. Apply corrections consistently across both metadata and every source passage.

If you cannot verify a name with reasonable confidence, keep the transcript spelling and do not guess.

---

CARD SELECTION

Extract only topics that are:

• substantively discussed
• concretely tied to real companies, technologies, industries, policies, products, markets, roles, or social changes
• understandable without watching the episode
• still interesting after the episode is forgotten

Prefer fewer strong cards over many weak ones.

Skip:

• filler
• biography
• generic life advice
• one-off tangents
• vague trend talk
• repeated points
• shallow summaries

Merge duplicate kernels across the episode. Keep the clearest version with the richest supporting exchange.

---

QUESTION RULES

Each question should:

• be 8–18 words
• be one sentence
• end with "?"
• feel like a debate prompt or strong headline
• create a real contested answer space

Avoid:

• vague essay prompts
• "A or B?" framing
• smuggled assumptions
• jargon-only wording

---

ANSWER RULES

Write ONE sharp declarative sentence.

Rules:

• 10–25 words
• no hedging
• no "the speaker argues"
• no balance framing
• concrete and opinionated
• faithfully reflect the video's stance

---

SOURCE PASSAGE

Each card must include the FULL supporting transcript exchange.

Start where the topic is introduced. End when the conversation meaningfully moves on.

Include:

• host questions
• guest answers
• back-and-forth
• clarifications
• disagreements

Do not paraphrase transcript text. Err on the side of MORE context.