You are building a short trailer for a podcast episode by picking 4 to 5
sentences from the transcript that, played back-to-back, work as ONE
coherent mini-narrative.

Hard requirements for what you pick:

  1. EXACTLY 4 OR 5 PICKS. Not 6, not 10, not 20. Be ruthless. If a
     candidate is only "good," cut it.

  2. EACH PICK IS A SELF-SUFFICIENT HOOK LINE. A stranger encountering
     it cold on social media would understand it AND want to keep
     watching. No mid-thought picks. No "...and that's why..." picks
     that depend on prior context.

  3. THE 4-5 PICKS, IN ORDER, FORM A COHERENT ARC. They should feel
     like they belong in the same trailer:
        opener         (the hook of the episode)
        development    (one or two reframes / sharp claims)
        payoff         (the line you want viewers leaving with)
     Picks from totally different parts of the conversation are fine
     ONLY IF they thematically connect.

  4. EACH PICK IS ONE COMPLETE SENTENCE (or at most two tightly-coupled
     sentences forming a single thought). Begin at the start of the
     sentence, end at the end of the sentence — never mid-clause.

Output a JSON array of EXACTLY 4 or 5 items, no prose around it. Each item:
  - "sentence": the verbatim sentence text from the transcript (so we
                can locate it precisely)
  - "start":    seconds (float) when the sentence begins
  - "end":      seconds (float) when the sentence ends — be generous;
                better to overshoot than to clip the last word

The transcript below has [MM:SS-MM:SS] timestamps per segment. Use those
to derive your "start" and "end" values.

Example output format:
[
  {"start": 109.82, "end": 113.40, "sentence": "Confidence is an output, not an input."},
  {"start": 681.50, "end": 685.20, "sentence": "Make anxiety sit in the back seat."}
]
