Lex Transcript Search

Lex Transcript Search

Lex Fridman publishes a human-written, timestamped transcript for every recent episode at https://lexfridman.com/<guest-slug>-transcript/. Each paragraph carries its speaker, a (hh:mm:ss) start time, and a link into the full YouTube episode. search.py fetches that page, parses it into segments, and greps them, so every timestamp you report comes straight from the page.

For non-Lex videos, or for a full "what's interesting in this" summary rather than finding a specific topic, use the youtube-digest skill instead.

Workflow

1. Identify the episode

• WebSearch Lex Fridman <guest> (or Lex Fridman podcast latest episode). You want the lexfridman.com/<slug>-transcript/ URL. The episode page lexfridman.com/<slug>/ also works, because the script follows its transcript link.
• Slugs are the guest's name, and repeat guests get -2, -3, and so on (dhh-2-transcript, sam-altman-2-transcript). If the user says "the latest DHH episode", make sure you have the highest-numbered one.
• Check that you have the right episode: the script prints the page title with the episode number (e.g. Lex Fridman Podcast #501). If that doesn't match the user's description (guest, rough date, topic), say so before answering.

2. Search: the literal word plus sensible variants

python3 "${CLAUDE_PLUGIN_ROOT}/skills/lex-transcript-search/search.py" <URL> <term> [<term> ...] [-c N] [--json]

• Terms are case-insensitive regexes, and all of them are searched at once. A plain stem matches every form: inferenc finds inference and inferencing. Use \bword\b when a short word would match inside other words (e.g. \bAI\b, \bcompute\b).
• For a concept (not an exact word), add synonyms, related names, and spelling variants in the same call. For "inference" that could be inferenc 'open.weight' Fireworks 'token'. For "security" it could be secur vulnerab exploit hack. Use . or [- ]? for hyphen/space variants (open.source), since the transcripts are typed by people and aren't consistent.
• -c 1 or -c 2 prints neighbouring segments, so you can tell what the conversation was about. --json gives structured output.
• Clips in the "Episode highlight" cold open at 0:00 are skipped by default, because they replay material from later in the episode. The script says how many it skipped. Use --include-highlights to show them.
• If there are no matches, the script says so and suggests variants. Try a shorter stem or synonyms before telling the user the topic never comes up.
• Pages are cached in ~/.cache/lex-transcript-search/. Use --refresh if a transcript was published or corrected recently.

3. Read around the hits

Run again with -c 2 on the interesting hits (or read the context field in --json) so that you understand the actual discussion. A keyword hit in passing is not the same as a real discussion. Tell those apart for the user.

4. Answer

For each relevant moment, give:

• Timestamp range and deep link, copied exactly from the script output: 02:40:06–02:40:43 → https://www.youtube.com/watch?v=NYFGCESmikA&t=9606s
• Speaker and a short verbatim quote (one or two sentences) containing the term.
• A plain-English explanation of what was being discussed and why it matters. Add a sentence of background if the user is unlikely to know the topic.

Rules:

• Never invent or estimate timestamps. Use only the ones the script printed. If the script can't find a video id, give the transcript timestamps without links.
• The timestamp is the start of the paragraph. A long paragraph can run about a minute, so the word may come 30–60s after the link starts. Say so when the quote is near the end of the paragraph. The -> end time is when the next segment begins.
• Group adjacent hits (back-to-back segments on the same thread) into one moment.
• Quote verbatim. The transcript is human-made and may have small errors, so don't "fix" names or model versions without saying that you did.

Limitations

• This only works for episodes that have a lexfridman.com transcript. Older or very new episodes may not have one yet. In that case fall back to youtube-digest.
• It matches text, not meaning, so the quality of your synonyms decides how much you find.
• The video id is the most-linked one on the page (the full episode). Clips or shorts on other channels have different timings.

Tests

cd "${CLAUDE_PLUGIN_ROOT}/skills/lex-transcript-search" && python3 -m unittest test_search.py (offline, synthetic HTML).