Meeting Transcript Action Item Extraction and Priority Prompt Template

This prompt is designed for knowledge workers and engineering teams who need to extract structured, prioritized action items from raw meeting transcripts. It uses few-shot examples to teach the model how to identify tasks, detect ownership, extract deadlines, and assign a priority score, rather than relying on verbose instruction sets. Use this when you need consistent, machine-readable outputs from messy, multi-speaker transcripts that will be ingested into project management tools, task queues, or automated follow-up systems. This is not a meeting summarization prompt; it focuses exclusively on action item extraction and triage.

The ideal input is a raw, diarized transcript where speakers are identified. The prompt expects a [TRANSCRIPT] placeholder containing the full meeting text. It works best when the transcript includes clear verbal commitments ("I will," "I'll take that"), explicit deadlines, and assigned owners. The output is a structured JSON array of action items, each with fields for task_description, owner, deadline, priority_score (1-5), and evidence_quote. This schema is enforced through few-shot examples, not just instructions, which makes the extraction more robust to variations in speaking style and transcript quality.

Do not use this prompt for generating meeting summaries, identifying decisions without assigned actions, or analyzing sentiment. It is not suitable for transcripts where speakers are not identified, as ownership detection relies on speaker labels. If your transcript lacks clear ownership signals, pair this prompt with a pre-processing step that attempts to resolve pronouns and implicit assignments. For high-stakes workflows where missed action items have compliance or legal consequences, always route outputs through a human review queue before ingestion into downstream systems. The prompt includes a confidence marker for each extracted item, which you should use to filter low-confidence extractions for manual verification.

This prompt template teaches the model to parse a raw meeting transcript and produce a structured list of action items. The few-shot examples are the primary teaching mechanism—they demonstrate extraction of implicit commitments, mapping of speakers to owners, deadline normalization, and priority assignment based on urgency signals in the conversation. The template is designed to be copied directly into your prompt management system or API call, with square-bracket placeholders you replace at runtime.

textCopy
You are an action item extraction system. Your task is to read a meeting transcript and output a JSON array of action items. Each action item must include: the task description, the owner (the person responsible), any mentioned deadline, and a priority level (High, Medium, Low).

Follow these rules:
- Extract only concrete commitments, tasks, or follow-ups. Do not invent tasks that were not discussed.
- If a task is assigned to a person by name, use that name as the owner. If the speaker volunteers, they are the owner. If no owner is clear, use "Unassigned".
- Normalize all deadlines to ISO 8601 dates (YYYY-MM-DD). If a relative date is given (e.g., "next Tuesday"), calculate the actual date assuming the meeting date is [MEETING_DATE]. If no deadline is mentioned, use null.
- Assign priority based on these signals: explicit urgency language ("ASAP", "critical", "blocking"), dependency on other tasks, or proximity to the deadline. Default to Medium if no signal is present.
- Output ONLY a valid JSON array. No other text.

[EXAMPLES]

---

Transcript:
Alice: Okay, let's wrap up. Bob, can you send the updated contract to the client by end of day Friday?
Bob: Yes, I'll get that out.
Alice: Great. And I need to review the Q3 budget before our next sync. That's critical—we can't move forward without it.
Charlie: I'll follow up with the vendor about the delayed shipment. No rush on that, maybe sometime next week.

Output:
[
  {
    "task": "Send updated contract to client",
    "owner": "Bob",
    "deadline": "2025-03-28",
    "priority": "High"
  },
  {
    "task": "Review Q3 budget",
    "owner": "Alice",
    "deadline": null,
    "priority": "High"
  },
  {
    "task": "Follow up with vendor about delayed shipment",
    "owner": "Charlie",
    "deadline": null,
    "priority": "Low"
  }
]

---

Transcript:
Dana: I'll draft the onboarding guide. Should be done by Wednesday.
Evan: I can review it once Dana's done. Let's aim to have it finalized by the 15th.
Dana: Sounds good. Also, someone needs to update the team wiki with the new process, but I'm not sure who has bandwidth.

Output:
[
  {
    "task": "Draft onboarding guide",
    "owner": "Dana",
    "deadline": "2025-03-26",
    "priority": "Medium"
  },
  {
    "task": "Review onboarding guide",
    "owner": "Evan",
    "deadline": "2025-04-15",
    "priority": "Medium"
  },
  {
    "task": "Update team wiki with new process",
    "owner": "Unassigned",
    "deadline": null,
    "priority": "Medium"
  }
]

---

Transcript:
[TRANSCRIPT]

Output:

To adapt this template, replace [MEETING_DATE] with the actual date of the meeting in YYYY-MM-DD format so relative deadlines can be resolved correctly. Replace [TRANSCRIPT] with the raw meeting text. The [EXAMPLES] block contains two few-shot demonstrations; you can add more examples that reflect your organization's typical task patterns, ownership conventions, and priority definitions. If your meetings use specific terminology for urgency (e.g., "P0", "drop everything"), include an example that maps that language to the priority field. Before deploying, run the prompt against a golden set of 10–20 annotated transcripts and measure extraction recall, owner accuracy, and priority agreement against human labels. For high-stakes workflows where missed action items have compliance or revenue impact, route outputs with "priority": "High" or "owner": "Unassigned" to a human review queue before ingestion into your task management system.

This prompt is designed to be the core extraction step within a larger meeting intelligence pipeline. It should not be called directly by an end-user in a chat interface. Instead, it belongs in a backend service that receives a transcript, enriches it with participant context, calls the model, and then validates and stores the structured output. The primary integration points are a transcript pre-processing stage, the LLM call itself, and a post-processing validation stage.

For implementation, wrap the prompt in a function that accepts a transcript string and an optional participants list. Before calling the model, inject the participant list into the [PARTICIPANT_CONTEXT] placeholder to improve ownership detection. Use a model with strong JSON mode and function-calling capabilities, such as gpt-4o or claude-3.5-sonnet, with response_format set to json_object and the JSON Schema provided in the [OUTPUT_SCHEMA] block. On the application side, implement a strict validation layer that checks the returned JSON against the schema. If validation fails, use a repair loop: feed the raw output and the validation error back to the model with a simplified repair prompt, such as 'Fix the JSON to match the schema. Return only the corrected JSON.' Limit this retry loop to a maximum of 2 attempts before logging the failure and alerting a human. Log every extraction, including the raw transcript hash, the model's raw response, and the validated output, for auditability and prompt evaluation.

For high-stakes environments where a missed action item has significant consequences, do not rely solely on this prompt. Implement a human review queue for all extracted items scored as priority: 'high' or where the confidence score is below 0.85. The application should route these items to a review interface where a human can confirm, reject, or modify the extraction. This human-in-the-loop step provides a safety net and generates a stream of correction data that can be used to fine-tune a model or improve the few-shot examples over time. Avoid deploying this as a fire-and-forget automation without these validation and review guardrails.