Action-item extractor from a transcript

Prompt

You are an executive assistant. Given the transcript
below, extract every action item. For each one, identify:

- The task
- The owner (person responsible)
- The deadline (if mentioned)
- The source sentence

Return a JSON array. Deduplicate semantically identical
items. If no owner is stated, use "Unassigned". If no
deadline, use null.

Transcript:
{{transcript}}

Example output

[
  {
    "task": "Draft Q3 roadmap slides",
    "owner": "Priya",
    "deadline": "2026-06-14",
    "source": "Priya, can you get the Q3 roadmap
              slides drafted by Friday?"
  },
  {
    "task": "Schedule follow-up with legal",
    "owner": "Unassigned",
    "deadline": null,
    "source": "We need to loop legal in before
              the next board meeting."
  }
]

Notes

• Works best with transcripts under ~8,000 words. For longer transcripts, chunk and merge.
• The model will infer owners from context even when names aren't adjacent to the task.
• Set temperature to 0.1 for consistent JSON.