podcast-intro

Long-form podcast intro (cold open)

The look is modeled on Greg Isenberg's long-form intros. What makes it engaging:

• Open on the strongest line, with no logo first.
• Cut every 3–6 s.
• Make roughly every other clip a cutaway: the speaker slides into a card on the left while a graphic lands on the right.
• Keep captions short: 1–3 words, with orange keywords.
• End on the title: a card that shows what's coming.

The pipeline is pipeline/build_episode.py, driven by work/<ep>/episode.json → intro. Rendering uses the Remotion IntroV2 composition (pipeline/remotion/src/IntroV2.tsx); graphics come from Graphics.tsx.

1. Mine the episode for clips (8–10 candidates, keep 7–9)

• Read the transcript (work/<ep>/words-*.json, or the Descript .txt) and list lines that stand alone in under 6 s.
• In order of preference:
  1. A hard number: "pay 27% of their employees to leave", "900% revenue retention".
  2. A bold claim or prediction: "tokens are going to seep into everything".
  3. A confession or relatable pain: "I'm so tired of paying subscriptions".
  4. A funny or emotional reaction: "I'd be token maxing".
  5. A quotable closer: "can't put a price on superintelligence".
• Spread across topics: about one clip per chapter, both hosts represented, at least one laugh.
• Order the reel:
  • Clip 1 is the single strongest hook, usually a number or a claim.
  • Alternate speakers and energy.
  • Put the most quotable line last, right before the title card.
• Skip:
  • lines needing context ("like I said", "that chart");
  • lines over ~6 s;
  • lines with crosstalk.
• Swearing: keep it only if the user is fine with an explicit intro, and say so in the report.

2. Configure episode.json → intro

"title": "FRINGE LINES",
"subtitle": "October 30, 2026",
"intro": {
  "style": "v2",
  "title": "Fringe Lines",
  "title_seconds": 4.5,
  "topics": ["The Morning Loop", "Microsoft headcount", "Build vs. buy", "Token limits", "Google vs. AI mode", "Anthropic's S-1", "AI spend"],
  "emphasis": ["27%", "leave", "900%", "retention", "sql", "agents", "tokens", "everything", "maxing"],
  "clips": [
    {"say": "microsoft just offered", "until": "employees to leave", "tag": "AI & jobs", "replace": {"27": "27%"},
     "graphic_on": "27%", "graphic": {"type": "stat", "kicker": "Microsoft buyout offer", "value": 27, "suffix": "%", "label": "of employees offered a payout to *leave*"}},
    {"say": "tokens are going to seep", "until": "like the internet", "tag": "Tokens everywhere"}
  ]
}

• say / until: the first and last few spoken words of the clip, matched case- and punctuation-insensitively. The match is the first occurrence in the episode, so make say distinctive (3–5 words). If the other speaker talks over the line, the matcher retries using only one speaker's words.
• pre: optional seconds of lead-in before say.
• tag: the topic label shown top-left with its number. Use 1–3 words.
• replace: caption fixes, e.g. numbers, "sequel"→"SQL", or "a": "" to hide a word.
• graphic + graphic_on: the cutaway lands about 150 ms before the spoken word graphic_on.
  • Optional graphic_off (a later word) returns to full-frame.
  • Without graphic_off, the cutaway holds to the end of the clip.
• Cutaway count: add cutaways to 3–5 clips. Pick the ones with a number, a named company or product, or a quotable line. Leave reaction and laugh clips full-frame.
• Graphic types (same as the editorial shorts):
  • stat: a counting number with kicker and label;
  • people: an icon grid;
  • lines: schematic trend lines;
  • quote: an ink card;
  • headline: a news-style card with a highlight sweep;
  • list: numbered rows.
  • Use *word* for orange italics.
• Honesty: on-screen numbers must be what was said. Charts are schematic, never invented data.
• topics: 5–7 short chapter names for the title card, in episode order.
• emphasis: 12–25 keywords drawn orange in captions: numbers, names, punchline words.
• title_seconds: 4–5 s. The title card is "Fringe Lines" in serif with an orange italic last word, the date, and an "On this episode" list.

3. Check the plan before rendering

~/.shorts-skill/bin/python pipeline/build_episode.py --ep <ep> --stage plan

• It prints each intro clip's tag, length and caption text.
• Fix any WARN intro clip not found: adjust the say/until wording to match the transcript exactly.
• Target total: clips of 35–50 s plus a 4.5 s title. Drop or trim clips over 6 s.
• Read every caption line for misheard words and add replace fixes.

4. Render

~/.shorts-skill/bin/python pipeline/build_episode.py --ep <ep> --stage graphics   # intro clips + IntroV2 + lower-thirds
~/.shorts-skill/bin/python pipeline/build_episode.py --ep <ep> --stage final      # join intro + body, chapters, loudness

• Add --blur to both commands for the background-blurred version. It needs the blurred cameras in ~/.cache/fringe-lines-video/blur/ (see pipeline/bg_blur.py) and writes <ep>-fringe-lines-blur.mp4 from a separate cache.
• The body doesn't need re-rendering for intro changes.
• The chapters file is regenerated, so all timestamps shift with the intro length.
• Sound effects (whoosh, pop, long whoosh into the title) are generated into the graphics folder and mixed by Remotion.

5. QA

F=out/episode/<ep>-fringe-lines.mp4
ffmpeg -v error -y -i $F -t 50 -vf "fps=1,scale=320:-1,tile=8x7" -frames:v 1 /tmp/intro.png

Check:

• Each cutaway's card is fully on screen, its numbers match what's said, and there's no stray *.
• Captions show orange keywords, and no caption overlaps the speaker card's bottom edge.
• The title card lists every topic in order.
• Video and audio durations match (ffprobe -show_entries stream=codec_type,duration).
• The episode sits at about −15.5 LUFS.
• The chapters file starts 00:00 Cold open, then the intro length, then "Catching up".

Report to the user:

• the clip list (tag + line);
• which clips have cutaways;
• the intro length;
• any explicit language;
• any numbers to verify.