Find viral moments in a transcript with an LLM

Both Gemma 4 runs read the same transcript of a 10-minute podcast. Workers AI prices Gemma 4 at $0.10 per million input tokens and $0.30 per million output tokens.

1. 1

   Number the transcript

   Give the model numbered lines with times, and ask for line numbers back. Numbers are easy to check, while times the model types itself can drift. A line ends at a full stop, a question mark or a pause over 0.7 s.

   import json
   
   words = json.load(open("words.json"))
   
   # One numbered line per sentence, or per stretch of speech that ends in a pause
   lines, current = [], []
   for i, word in enumerate(words):
       current.append(word)
       next_start = words[i + 1]["start"] if i + 1 < len(words) else None
       if word["word"][-1] in ".?!" or next_start is None or next_start - word["end"] > 0.7:
           lines.append({
               "id": len(lines),
               "start": current[0]["start"],
               "end": word["end"],
               "text": " ".join(w["word"] for w in current),
           })
           current = []
   
   json.dump(lines, open("lines.json", "w"))
   print(len(lines), "lines")
   for line in lines[60:63]:
       print(f'[{line["id"]}] {line["start"]:.1f}-{line["end"]:.1f}: {line["text"]}')
   
   # $ python3 lines.py
   # 158 lines
   # [60] 236.7-237.9: Now the science begins.
   # [61] 237.9-246.9: The scary part of getting off the earth, traveling, you know, 300 million miles and landing on the surface is over, now the science can begin.
   # [62] 247.9-260.0: You know, one of the things I think of when they do the EDL, when it lands, is that I think, I don't know if it was during Curiosity or one of the recent missions where, I believe there was a choreographed dance move.
   

2. 2

   Ask for hook, complete thought and payoff

   The prompt names the 3 things a clip needs, the length in seconds and the rules, and a JSON schema makes the answer parse every time. The model also returns its hook and payoff lines, so you can see why it picked each clip.

   import json
   from workers_ai import run
   
   lines = json.load(open("lines.json"))
   transcript = "\n".join(f'[{l["id"]}] {l["start"]:.1f}-{l["end"]:.1f}: {l["text"]}' for l in lines)
   
   prompt = f"""Below is a podcast transcript. Each line is [id] start-end seconds: text.
   
   Pick the 3 best moments to post as standalone clips on TikTok, Reels and Shorts.
   A good clip has all three of these:
   1. Hook: the first line makes a stranger stop scrolling (a question, a bold claim, a surprise).
   2. Complete thought: it makes sense without the rest of the episode.
   3. Payoff: it ends on the answer, the punchline or the conclusion, never mid-thought.
   
   Rules:
   - A clip is a run of whole lines, from first_line to last_line.
   - A clip lasts 20 to 60 seconds: last line's end minus first line's start.
   - Clips must not overlap.
   - Skip intros, outros, thank-yous and plugs.
   - Score each clip 0 to 100 for how well it works on its own.
   
   Transcript:
   {transcript}"""
   
   schema = {
       "type": "object",
       "properties": {
           "clips": {
               "type": "array",
               "items": {
                   "type": "object",
                   "properties": {
                       "first_line": {"type": "integer"},
                       "last_line": {"type": "integer"},
                       "hook": {"type": "string"},
                       "payoff": {"type": "string"},
                       "score": {"type": "integer"},
                   },
                   "required": ["first_line", "last_line", "hook", "payoff", "score"],
               },
           }
       },
       "required": ["clips"],
   }
   
   result = run("@cf/google/gemma-4-26b-a4b-it", {
       "messages": [{"role": "user", "content": prompt}],
       "response_format": {"type": "json_schema", "json_schema": {"name": "clips", "schema": schema, "strict": True}},
   })
   
   picks = json.loads(result["choices"][0]["message"]["content"])["clips"]
   json.dump(picks, open("picks.json", "w"), indent=2)
   print(json.dumps(picks, indent=2))
   print(result["usage"]["prompt_tokens"], "prompt tokens,", result["usage"]["completion_tokens"], "completion tokens")
   
   # $ time python3 pick.py
   # [
   #   {
   #     "first_line": 22,
   #     "last_line": 34,
   #     "hook": "It's a big deal. It is. So much so they call it the seven minutes of terror.",
   #     "payoff": "I mean, all of these things have to happen right.",
   #     "score": 95
   #   },
   #   {
   #     "first_line": 53,
   #     "last_line": 60,
   #     "hook": "Why do you see people jumping for joy and hugging each other and crying?",
   #     "payoff": "Now the science begins.",
   #     "score": 90
   #   },
   #   {
   #     "first_line": 113,
   #     "last_line": 119,
   #     "hook": "I mean, what would be a sign of ancient life?",
   #     "payoff": "So there are these biosignatures that really hint more strongly at the fact that it comes from something that was living.",
   #     "score": 88
   #   }
   # ]
   # 5194 prompt tokens, 8186 completion tokens
   #
   # real	1m35.817s
   # user	0m0.090s
   # sys	0m0.023s
   

3. 3

   Check every pick

   Don't cut what the model says without checking it. This script drops picks with missing line numbers, picks outside 20 to 60 s and picks that overlap a better one. All 3 of ours passed.

   import json
   
   lines = {l["id"]: l for l in json.load(open("lines.json"))}
   picks = json.load(open("picks.json"))
   
   # Turn line ids into times, then drop anything the rules forbid
   kept = []
   for pick in sorted(picks, key=lambda p: -p["score"]):
       first, last = lines.get(pick["first_line"]), lines.get(pick["last_line"])
       if first is None or last is None or last["id"] < first["id"]:
           print("bad line ids:", pick)
           continue
       start, end = first["start"], last["end"]
       if not 20 <= end - start <= 60:
           print(f"wrong length ({end - start:.1f} s):", pick["hook"])
           continue
       if any(start < k["end"] and k["start"] < end for k in kept):
           print("overlaps a better clip:", pick["hook"])
           continue
       kept.append({"start": start, "end": end, "score": pick["score"], "hook": pick["hook"]})
   
   for clip in kept:
       print(f'{clip["start"]:.1f}-{clip["end"]:.1f} s ({clip["end"] - clip["start"]:.1f} s) score {clip["score"]}: {clip["hook"]}')
   json.dump(kept, open("clips.json", "w"), indent=2)
   
   # $ python3 check.py
   # 76.9-123.6 s (46.7 s) score 95: It's a big deal. It is. So much so they call it the seven minutes of terror.
   # 187.8-237.9 s (50.1 s) score 90: Why do you see people jumping for joy and hugging each other and crying?
   # 421.7-462.1 s (40.4 s) score 88: I mean, what would be a sign of ancient life?
   

4. 4

   Keep the model's thinking on

   Turning thinking off brought the answer back in about 6 seconds instead of 96, but none of its 3 picks passed: one ran backwards from line 79 to line 34, and two named lines past the end of the transcript.

   # pick_fast.py is pick.py with one more field in the request,
   # and it writes picks_fast.json:
   #     "chat_template_kwargs": {"enable_thinking": False},
   # check_fast.py is check.py reading picks_fast.json
   
   $ time python3 pick_fast.py
   [
     {
       "first_line": 79,
       "last_line": 34,
       "hook": "So much so they call it the seven minutes of terror.",
       "payoff": "I mean, all of these things have to happen right.",
       "score": 95
     },
     {
       "first_line": 337,
       "last_line": 394,
       "hook": "Yeah, so even down to the drill, if you look at prior rovers, like the Curiosity rover, it has a drill that's meant to grind the soil and the dirt and the rocks into a fine powder.",
       "payoff": "It's about 10 centimeters long.",
       "score": 88
     },
     {
       "first_line": 549,
       "last_line": 570,
       "hook": "Do you think there might be ancient life on Mars?",
       "payoff": "It doesn't matter what I think. Let's see what the data says.",
       "score": 92
     }
   ]
   5191 prompt tokens, 272 completion tokens
   
   real	0m5.937s
   user	0m0.078s
   sys	0m0.024s
   
   $ python3 check_fast.py
   bad line ids: {'first_line': 79, 'last_line': 34, 'hook': 'So much so they call it the seven minutes of terror.', 'payoff': 'I mean, all of these things have to happen right.', 'score': 95}
   bad line ids: {'first_line': 549, 'last_line': 570, 'hook': 'Do you think there might be ancient life on Mars?', 'payoff': "It doesn't matter what I think. Let's see what the data says.", 'score': 92}
   bad line ids: {'first_line': 337, 'last_line': 394, 'hook': "Yeah, so even down to the drill, if you look at prior rovers, like the Curiosity rover, it has a drill that's meant to grind the soil and the dirt and the rocks into a fine powder.", 'payoff': "It's about 10 centimeters long.", 'score': 88}