An AI prompt to extract action items, owners and due dates from a meeting transcript

To extract action items, owners and due dates from a meeting transcript with an AI prompt, put the transcript inside clear tags and give the model firm rules: quote the line each item comes from, name an owner only when the transcript names one, keep each due date in the words people said, and write unknown when either is missing. Questions nobody answered go in a list of their own. The example prompt below follows those rules, and it is plain text you can paste into any AI chat.

The techniques we credit to Anthropic, OpenAI and Google come from the prompting guides they publish for their own models, which we read on October 16, 2026. The prompt, the transcript and the output on this page are made-up examples written for this guide. We ran no hands-on tests and took no screenshots. Near the end, you will see how Barside, the free desktop app we make for Windows, Linux and Mac, runs the same checks on a transcript in a note.

The prompt that pulls action items, owners and due dates from a meeting transcript

This AI prompt extracts action items, owners and due dates from a meeting transcript in one pass. Paste your transcript with a <transcript> tag on the line before it and a </transcript> tag on the line after it, then paste the instructions. Keep the speaker names your transcript tool wrote, and keep labels such as Speaker 2 when you do not know who spoke.

Example prompt (made up for this guide). Copy the paragraph and the eight numbered rules below:

The meeting transcript is between the <transcript> tags. I will send the action items you find to everyone who attended, so a wrong owner or a guessed date would send someone after the wrong task. Follow these rules:

1. Use only what the transcript says. Do not use general knowledge, and do not guess what people meant.
2. Find every action item: a task someone agreed to do, offered to do, or accepted when asked, and a task people said must be done even if no one took it.
3. For each action item, give the task as a short phrase that starts with a verb, the owner, the due date, and the exact line or lines it came from, each copied word for word with its timestamp.
4. The owner is the name the transcript uses for the person who took the task. If the person appears only as a label, such as Speaker 2, write that label. If no one took the task, write unknown.
5. Write the due date in the speaker's own words, such as "by Thursday" or "next week". If no deadline was said, write unknown. Never turn a spoken deadline into a calendar date, and never add a date nobody said.
6. A question nobody answered, or a topic people agreed to decide later, is not an action item. List it under Open questions, with its quote.
7. Before you answer, check every item against the transcript. If you cannot find its quote, remove the item.
8. Answer with a table with the columns Task, Owner, Due (as said) and Quote, then the Open questions list. If you find no action items, write "No action items found."

You can rename the columns, add a Decisions list or change the opening sentence to fit your team. Keep rules 4, 5 and 7, because they are what stop a guess from looking like a fact in the finished list.

A made-up transcript, and the output the prompt should give

A short sample shows the rules at work. Example transcript (made up): a weekly sync of a small team launching a newsletter, with Priya, the editor, Marco, the designer, and Dana, the developer.

• [00:40] Priya: Launch is still Friday. Marco, where are the new signup page designs?
• [00:52] Marco: Almost done. I'll send the mockups by Thursday.
• [01:30] Dana: The welcome email still breaks on phones. I'll fix the template.
• [02:05] Priya: Someone has to confirm the sponsor slot before Friday.
• [02:40] Priya: Dana, can you also set up the subscriber import next week?
• [02:46] Dana: Sure, I'll take that.
• [03:15] Marco: Are we still offering a yearly plan at launch?
• [03:22] Priya: Let's leave that for next week's call.

Example output (made up, written to follow the prompt's rules):

Open questions:

• Are we still offering a yearly plan at launch? [03:15] Marco: "Are we still offering a yearly plan at launch?" and [03:22] Priya: "Let's leave that for next week's call."

Look at what the rules did with this sample. The sponsor slot has a deadline and no owner, so its Owner reads unknown, and that is the first thing to settle in the follow-up email. Dana's template fix has an owner and no date. The import stays "next week", because only the team knows which week Priya meant. The yearly plan question lands under Open questions instead of becoming somebody's task, and "Launch is still Friday" is a decision, which this prompt does not ask for.

Why each rule is in the prompt

Most rules map to advice in the vendors' own guides. The rest, such as keeping deadlines in the speaker's words and listing open questions apart, are our own choices for a follow-up list. Knowing which is which helps you adjust the prompt without losing what makes it reliable.

• Tags around the transcript. Anthropic's prompting best practices say XML tags "help Claude parse complex prompts unambiguously, especially when your prompt mixes instructions, context, examples, and variable inputs." OpenAI's prompt engineering guide also suggests Markdown and XML to mark the parts of a prompt, with XML tags showing where a reference document begins and ends.
• A quote for every item. Anthropic's guide to reducing hallucinations suggests asking for word-for-word quotes first on long documents. It also suggests having the model find a supporting quote for each claim and retract any claim it cannot support. Rule 7 is that second step.
• Unknown instead of a guess. The same Anthropic guide says to give Claude explicit permission to admit uncertainty. Google's prompt design strategies for Gemini include a sample system instruction that tells the model to say the information is not available when the exact answer is not written in the context.
• Only the transcript. Anthropic lists external knowledge restriction: tell the model to use only the documents you provide, and not its general knowledge. Rule 1 does this.
• The reason in the opening paragraph. Anthropic's best practices recommend explaining why an instruction matters, because that context helps the model understand your goal. The line about sending the list to everyone who attended gives that reason.
• Numbered rules and a fixed format. Anthropic suggests numbered steps when the order or completeness of steps matters, and being specific about the output format. Google's guide says you can specify constraints and the format of the response.

The guides differ on one point: where the transcript goes. Anthropic's long context tips put long documents near the top, above the instructions, for inputs of 20,000 tokens or more. OpenAI's guide says context is usually best near the end of the prompt, because you may include different context for different requests. Follow the guide of the AI you use; the tags mark where the transcript starts and ends in either order.

Ask for JSON when the list feeds another tool

A structured extraction of action items, owners and due dates from an AI meeting transcript usually means JSON in place of a table, so a spreadsheet, a task app or a script can read the result. Replace rule 8 with an example rule like this one: answer only with JSON, an object with an action_items array and an open_questions array. Each action item has four keys:

• task: the short phrase that starts with a verb.
• owner: the name or label from the transcript, or the string "unknown".
• due_as_said: the deadline in the speaker's words, or "unknown".
• quote: the line the item came from, with its timestamp.

Each open question gets a question and a quote. A fixed string such as "unknown" is easier to filter in a spreadsheet than an empty cell, and the key name due_as_said reminds everyone who reads the data that it is not a calendar date.

If you call a model from code, the APIs can hold the answer to a shape. Anthropic's structured outputs page says the feature makes Claude's responses follow a JSON schema you send, set through output_config.format, so the result can be parsed. The page also lists cases where the output may not match the schema, such as a refusal or an answer cut off at the token limit. OpenAI's prompt engineering guide names the same idea Structured Outputs, which makes the JSON a model returns conform to a JSON schema.

A schema controls the shape of the answer. It cannot tell whether Dana really took the import, so keep the quote key and check it the same way you would check a table.

Best practices for action items, owners and due dates

The best practices for turning an AI meeting transcript into action items with owners and due dates start before anyone opens a chat. A transcript can only give the AI what people said out loud, and the AI's answer still needs a human check.

• Say names and dates out loud in the meeting. "Dana, can you take the import by Wednesday?" gives the AI an owner and a deadline it can quote. "Can someone look at that?" gives it neither.
• Fix speaker labels you are sure of before you paste. If your tool wrote Speaker 2 and you know it was Marco, change it, so the Owner column can name him. Leave the labels you are unsure of.
• Turn spoken deadlines into dates yourself. Keep "by Thursday" in the list, then pick the real date when you send the follow-up, since you know which Thursday was meant.
• Check the quotes. Search the transcript for each quote. If it is not there word for word, delete the item. Anthropic's hallucination guide says its techniques reduce made-up answers without removing them, and adds: "Always validate critical information, especially for high-stakes decisions."
• Run a second pass on long or messy calls. Anthropic describes a self-correction chain: generate a draft, have the model review it against criteria, then refine it, each as a separate call. Google's guide suggests making each step a prompt and chaining the prompts. A second prompt can hand back the transcript and the list and ask the model to confirm each quote and owner, then remove or fix every item that fails.
• Compare two runs. Anthropic's best-of-N verification runs the same prompt several times and compares the outputs, since differences between them can point to a hallucination. An item that shows up in one run only deserves a second look.
• Add examples to steer the format. Anthropic's best practices call examples one of the most reliable ways to steer format, suggest 3 to 5 of them for the best results, and suggest wrapping them in <example> tags so Claude can tell them apart from instructions. Google's guide recommends including few-shot examples, and OpenAI's suggests showing a diverse range of inputs with the outputs you want. Include an item with an unknown owner, so the model sees that unknown is a normal answer, and keep the tags so it does not take the example's tasks for tasks from your meeting.
• Mind where the transcript goes. Pasting a transcript into an AI chat sends it to that AI company, whose data policy applies. Read our guide to AI note apps and client confidentiality before you paste a client call, and tell people before you record one.

How Barside does this from a transcript in a note

Barside's Meeting follow-through runs the same kind of extraction on any note that holds a transcript or meeting notes, and it checks the AI's answer against the transcript before you see it. It uses the AI you connect in Settings → AI: Claude, ChatGPT, Gemini or OpenRouter. Barside has no AI of its own and does not charge for AI; that company bills you for what you use. Read more about bringing your own AI key.

1. Get the transcript into a note. Paste one from another meeting app, record a call with Transcribe a call, hold a Barside video meeting with captions on, or import Granola meetings through Settings → Import notes. Transcribe a call records your microphone and your computer's sound and turns them into text with Whisper on your computer. See transcribe calls on your computer.
2. Open the note and click Meeting follow-through next to its title. You can also open Library & briefing and choose the Meeting follow-through tab.
3. Pick the note under Transcript note and press Analyze meeting. Long calls are read in parts, and Stop cancels at any time.
4. Review the results. You get a Meeting summary, Decisions with the quote behind each one, Proposed tasks with an Owner, the deadline words people used and the quote, and a Follow-up draft.
5. Choose what to keep. Tick Keep this task or press Select all tasks, fix the Task text and the Owner, and pick a Reminder time for tasks that need an alarm, or leave it blank.
6. Press Save recap + N selected tasks. Barside makes a recap note with the draft marked "not sent", and one note per kept task, with its owner, the meeting it came from and the quote.

The checks match the rules of the prompt above. Every decision and task needs a quote that appears in the transcript, and anything without one is dropped. If the owner's name is not in the transcript, the task says Unassigned. Deadlines stay in people's words, and no reminder is set unless you choose a time. The AI is told not to turn a question into a decision, and not to invent commitments, owners, dates or quotes. For a long call, it is asked for at most 12 decisions and 12 tasks per part, and an item that repeats with the same wording or quote is kept once.

Saved tasks stay visible. Each one appears in the daily briefing under Unfinished commitments until you press Mark complete, and the meeting's decisions show under Recent decisions for 14 days, as long as the transcript note stays unchanged.

For a follow-up after every call, open Agents & approvals, choose the Meeting tab and tick After each call transcript, prepare a follow-up. It is off by default. The agent proposes a Follow-up checklist, reminders for up to six tasks with a clear deadline and a recap email, and each proposal waits in your approval inbox. When people named only a day, such as "by Friday", the proposed reminder is 9:00 AM that day, and a deadline that is unclear or already past gets no reminder. Read about AI agents that ask before they act.

If you prefer to work in Claude Desktop, ChatGPT or another MCP app, connect it through AI connections and pick the Meeting follow-up prompt. Your AI app lists the decisions and action items with owners and dates, and the prompt tells it to ask you before it adds anything as a task. The connection reaches only the workspaces you share with it.

Know the limits before you rely on it. On calls in other apps, Transcribe a call labels lines You or Others, so everyone but you is Others, and an owner needs to be named out loud to appear. It transcribes calls as it records them and does not transcribe an audio file you already have. Notes excluded from Library AI cannot be analyzed until you change their Sources setting.

Privacy: where the transcript goes

A prompt in an AI chat sends the whole transcript to that AI company. Barside works the same way at the AI step and keeps the rest on your computer:

• Meeting follow-through sends the transcript text to the AI company you connected in Settings → AI, and that company's data policy applies.
• With Transcribe a call, the audio stays on your computer, where speech is turned into text.
• The review, the recap and the task notes are saved in your library and its ZIP backups. Nothing is emailed or posted for you.

Read how Library and meeting analysis use your data and how recordings and transcripts work.

How a prompt compares with a meeting app

A hand-written prompt costs nothing to try, works with any transcript you can copy, and gives you full control over the rules. The checking is yours, though: the model can still slip, and nothing stops a missing quote unless you look. Barside's Meeting follow-through runs the quote check for you, saves tasks with reminders and keeps commitments in your daily briefing, with the AI account you connect. If you want a meeting tool that also joins or records calls, our guide to the best AI meeting assistants compares nine of them from their own pages, and the guide to AI note takers without a bot covers the ones that record from your computer. For a manual routine, the meeting notes template has a Decisions section and an Action items checklist with the owner's name first.

Works well with