feat(youtube): extract transcript highlights like Reddit comment gems

Add extract_transcript_highlights() that scores sentences by specificity
(numbers, proper nouns, topic relevance) and filters YouTube filler
(subscribe, welcome back, etc). Top 5 highlights shown as structured
bullets in compact output. Full transcript moved to collapsible <details>
block so the LLM reads highlights first, full text on demand.

SKILL.md updated to instruct the judge agent to quote highlights
directly in synthesis, same as Reddit top comments.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

SKILL.md

@@ -240,7 +240,7 @@ The script will automatically:
 
 **Read the ENTIRE output.** It contains EIGHT data sections in this order: Reddit items, X items, YouTube items, TikTok items, Instagram Reels items, Hacker News items, Polymarket items, and WebSearch items. If you miss sections, you will produce incomplete stats.
 
-**YouTube items in the output look like:** `**{video_id}** (score:N) {channel_name} [N views, N likes]` followed by a title, URL, and optional transcript snippet. Count them and include them in your synthesis and stats block.
+**YouTube items in the output look like:** `**{video_id}** (score:N) {channel_name} [N views, N likes]` followed by a title, URL, **transcript highlights** (pre-extracted quotable excerpts from the video), and an optional full transcript in a collapsible section. **Quote the highlights directly in your synthesis** - they are the YouTube equivalent of Reddit top comments. Attribute quotes to the channel name. Count them and include them in your synthesis and stats block.
 
 **TikTok items in the output look like:** `**{TK_id}** (score:N) @{creator} [N views, N likes]` followed by a caption, URL, hashtags, and optional caption snippet. Count them and include them in your synthesis and stats block.
 
@@ -303,9 +303,10 @@ The Judge Agent must:
 3. Weight TikTok sources HIGH (they have views, likes, and caption content — viral signal)
 4. Weight WebSearch sources LOWER (no engagement data)
 5. **For Reddit: Pay special attention to top comments** — they often contain the wittiest, most insightful, or funniest take. When a top comment has high upvotes (shown as `💬 Top comment (N upvotes)`), quote it directly in your synthesis. Reddit's value is in the comments.
-6. Identify patterns that appear across ALL sources (strongest signals)
+6. **For YouTube: Quote transcript highlights directly in your synthesis.** These are pre-extracted key moments from the video - treat them like Reddit top comments. Attribute to the channel name and include the actual quote. YouTube's value is in what creators SAY, not just their view counts.
-7. Note any contradictions between sources
+7. Identify patterns that appear across ALL sources (strongest signals)
-8. Extract the top 3-5 actionable insights
+8. Note any contradictions between sources
+9. Extract the top 3-5 actionable insights
 
 7. **Cross-platform signals are the strongest evidence.** When items have `[also on: Reddit, HN]` or similar tags, it means the same story appears across multiple platforms. Lead with these cross-platform findings - they're the most important signals in the research.
 

scripts/lib/normalize.py

@@ -193,6 +193,7 @@ def normalize_youtube_items(
             date_confidence="high",
             engagement=engagement,
             transcript_snippet=item.get("transcript_snippet", ""),
+            transcript_highlights=item.get("transcript_highlights", []),
             relevance=item.get("relevance", 0.7),
             why_relevant=item.get("why_relevant", ""),
         ))

scripts/lib/render.py

@@ -257,8 +257,15 @@ def render_compact(report: schema.Report, limit: int = 15, missing_keys: str = "
             lines.append(f"**{item.id}** (score:{item.score}) {item.channel_name}{date_str}{eng_str}{_xref_tag(item)}")
             lines.append(f"  {item.title}")
             lines.append(f"  {item.url}")
+            if item.transcript_highlights:
+                lines.append("  Highlights:")
+                for hl in item.transcript_highlights[:5]:
+                    lines.append(f'    - "{hl}"')
             if item.transcript_snippet:
-                lines.append(f"  Transcript: {item.transcript_snippet}")
+                word_count = len(item.transcript_snippet.split())
+                lines.append(f"  <details><summary>Full transcript ({word_count} words)</summary>")
+                lines.append(f"  {item.transcript_snippet}")
+                lines.append("  </details>")
             lines.append(f"  *{item.why_relevant}*")
             lines.append("")
 

scripts/lib/schema.py

@@ -210,6 +210,7 @@ class YouTubeItem:
     date_confidence: str = "high"  # YouTube dates are always reliable
     engagement: Optional[Engagement] = None
     transcript_snippet: str = ""
+    transcript_highlights: List[str] = field(default_factory=list)
     relevance: float = 0.7
     why_relevant: str = ""
     subs: SubScores = field(default_factory=SubScores)
@@ -226,6 +227,7 @@ class YouTubeItem:
             'date_confidence': self.date_confidence,
             'engagement': self.engagement.to_dict() if self.engagement else None,
             'transcript_snippet': self.transcript_snippet,
+            'transcript_highlights': self.transcript_highlights,
             'relevance': self.relevance,
             'why_relevant': self.why_relevant,
             'subs': self.subs.to_dict(),
@@ -655,6 +657,7 @@ class Report:
                 date_confidence=y.get('date_confidence', 'high'),
                 engagement=eng,
                 transcript_snippet=y.get('transcript_snippet', ''),
+                transcript_highlights=y.get('transcript_highlights', []),
                 relevance=y.get('relevance', 0.7),
                 why_relevant=y.get('why_relevant', ''),
                 subs=subs,

scripts/lib/youtube_yt.py

@@ -38,6 +38,52 @@ TRANSCRIPT_MAX_WORDS = 5000
 from .relevance import token_overlap_relevance as _compute_relevance
 
 
+def extract_transcript_highlights(transcript: str, topic: str, limit: int = 5) -> List[str]:
+    """Extract quotable highlights from a YouTube transcript.
+
+    Similar to reddit_enrich.extract_comment_insights() but for
+    continuous speech-to-text rather than threaded comments.
+    """
+    if not transcript:
+        return []
+
+    sentences = re.split(r'(?<=[.!?])\s+', transcript)
+
+    filler = [
+        r"^(hey |hi |what's up|welcome back|in today's video|don't forget to)",
+        r"(subscribe|like and comment|hit the bell|check out the link|down below)",
+        r"^(so |and |but |okay |alright |um |uh )",
+        r"(thanks for watching|see you (next|in the)|bye)",
+    ]
+
+    topic_words = [w.lower() for w in topic.lower().split() if len(w) > 2]
+
+    candidates = []
+    for sent in sentences:
+        sent = sent.strip()
+        words = sent.split()
+        if len(words) < 8 or len(words) > 50:
+            continue
+        if any(re.search(p, sent, re.IGNORECASE) for p in filler):
+            continue
+
+        score = 0
+        if re.search(r'\d', sent):
+            score += 2
+        if re.search(r'[A-Z][a-z]+', sent):
+            score += 1
+        if '?' in sent:
+            score += 1
+        sent_lower = sent.lower()
+        if any(w in sent_lower for w in topic_words):
+            score += 2
+
+        candidates.append((score, sent))
+
+    candidates.sort(key=lambda x: -x[0])
+    return [sent for _, sent in candidates[:limit]]
+
+
 def _log(msg: str):
     """Log to stderr."""
     sys.stderr.write(f"[YouTube] {msg}\n")
@@ -345,11 +391,15 @@ def search_and_transcribe(
     top_ids = [item["video_id"] for item in items[:transcript_limit]]
     transcripts = fetch_transcripts_parallel(top_ids)
 
-    # Step 3: Attach transcripts to items
+    # Step 3: Attach transcripts and extract highlights
+    core_topic = _extract_core_subject(topic)
     for item in items:
         vid = item["video_id"]
         transcript = transcripts.get(vid)
         item["transcript_snippet"] = transcript or ""
+        item["transcript_highlights"] = extract_transcript_highlights(
+            transcript or "", core_topic,
+        )
 
     return {"items": items}
 

tests/test_youtube_yt.py

@@ -45,5 +45,36 @@ class TestYtDlpFlags(unittest.TestCase):
         self.assertIn("--no-cookies-from-browser", cmd)
 
 
+class TestExtractTranscriptHighlights(unittest.TestCase):
+    def test_extracts_specific_sentences(self):
+        transcript = (
+            "Hey guys welcome back to the channel. "
+            "In today's video we're looking at something special. "
+            "The Lego Bugatti Chiron took 13,438 hours to build with over 1 million pieces. "
+            "Don't forget to subscribe and hit the bell. "
+            "The tolerance on each brick is 0.002 millimeters which is insane for injection molding. "
+            "So yeah that's pretty cool. "
+            "Thanks for watching see you next time."
+        )
+        highlights = youtube_yt.extract_transcript_highlights(transcript, "Lego")
+        self.assertTrue(len(highlights) > 0)
+        # Should pick the sentences with numbers and topic relevance, not filler
+        joined = " ".join(highlights)
+        self.assertIn("13,438", joined)
+        self.assertNotIn("subscribe", joined)
+        self.assertNotIn("welcome back", joined)
+
+    def test_empty_transcript(self):
+        self.assertEqual(youtube_yt.extract_transcript_highlights("", "test"), [])
+
+    def test_respects_limit(self):
+        sentences = ". ".join(
+            f"The model {i} has {i * 100} parameters and runs at {i * 10} tokens per second"
+            for i in range(20)
+        ) + "."
+        highlights = youtube_yt.extract_transcript_highlights(sentences, "model", limit=3)
+        self.assertEqual(len(highlights), 3)
+
+
 if __name__ == "__main__":
     unittest.main()