Internal Comms & Meetings

How to write meeting notes with AI

Internal Comms & Meetings3 AI tools7 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

Meeting notes are one of those tasks that sounds simple until you're the one doing it. Someone has to track who said what, capture the decisions that got made, log the action items and who owns them, and produce something useful enough that people actually read it afterward. On small teams, that job usually falls to whoever's running the meeting — which is often you, which means you're supposed to be leading the conversation and typing at the same time.

The reason people reach for AI here is obvious: most meeting notes follow a predictable structure. There's a summary, a list of decisions, a list of action items with owners, and maybe a next-steps section. That structure is something a language model can reliably fill in if you give it good raw material — a transcript, rough notes, or even a voice recording turned to text. The pattern-matching is exactly what LLMs are built for.

ChatGPT, Claude, and Gemini can all help meaningfully with this today. Paste in a transcript and ask for a structured summary with action items, and you'll get something usable within seconds. The output quality depends heavily on the quality of your transcript and the specificity of your prompt, but for a one-off meeting, a general-purpose LLM does a respectable job. The friction shows up in the workflow around it — getting the transcript in, deciding on a consistent format, and making sure the output goes somewhere people actually check.

Skip the manual work → See the Starch version

Internal Comms & Meetings3 AI tools7 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Record your meeting using Zoom, Google Meet, or any call platform that provides auto-transcription. Download the transcript as a text file when the call ends — most platforms export this directly.

2 Open ChatGPT, Claude, or Gemini and paste the full transcript into the chat window. If the transcript is long, Claude handles larger inputs more reliably than ChatGPT's free tier.

3 Prompt the model to produce structured meeting notes: a 3-5 sentence summary, a bulleted list of key decisions, and a table of action items with the owner's name and a due date extracted from the conversation.

4 Review the output carefully. LLMs occasionally misattribute action items or hallucinate a decision that was actually still under debate — read through and correct before sending to your team.

5 Copy the cleaned output and paste it into whatever tool your team uses — a Notion page, a Google Doc, a Slack message, or an email thread. This step is entirely manual; there is no automatic routing.

6 Save the prompt you used so you can reuse it next week. Without this, you'll rewrite or tweak the prompt from memory each time, and the format will drift across meetings.

7 If the transcript includes names the model doesn't know (e.g., 'J.T.' or 'the investor'), add a brief glossary at the top of your prompt so attributions land correctly in the output.

Prompts you can copy

Here is the transcript from our team sync. Write structured meeting notes: a 4-sentence summary, a bulleted list of decisions made, and a table of action items with owner name and due date. Transcript: [paste here]

Summarize this investor call transcript. Focus on: what the investor asked, what commitments we made, and any open questions we need to follow up on before next week. Keep the summary under 200 words.

From the transcript below, extract every action item. Format as a table with columns: Task, Owner, Due Date. If no due date was mentioned, write 'not specified.' Transcript: [paste here]

Rewrite these rough bullet-point notes from a 45-minute product meeting into clean meeting notes. Structure: summary, decisions, action items, open questions. Keep the tone professional but direct. Notes: [paste here]

This transcript is from a customer call. Pull out: (1) the customer's main complaints, (2) any feature requests they mentioned by name, (3) any follow-up we promised. Format as three separate lists.

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No connection to your calendar or meeting tool — you have to manually find the transcript, download it, and paste it in every single time.

Action items live in the chat window and nowhere else. Getting them into a task manager, Slack, or email requires a second round of copy-pasting by hand.

Output format drifts. The table structure you carefully prompted last Tuesday looks different from what you get next Tuesday unless you save and reuse the exact same prompt — which most people don't.

Large transcripts get truncated or produce degraded output. A two-hour all-hands transcript may exceed what the model handles cleanly, especially on free or mid-tier plans.

Nothing is searchable across meetings. Six weeks later when someone asks 'what did we decide about the pricing change?', there's no place to search — just a trail of chat sessions or a folder of Google Docs no one organized.

Attributing action items to the right people requires the model to recognize names from transcript shorthand, which it sometimes gets wrong — and catching those errors requires reading the full output carefully every time.

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system — for meeting notes, that means an agent builds and runs a persistent app connected to your actual calendar, transcripts, and task system, so structured notes, action items, and a searchable meeting archive happen automatically instead of through a prompt you re-run by hand each time.

The Meeting Notes starter app transcribes in real time, generates a structured summary after every call, and extracts action items with owners — so you're not pasting transcripts into a chat window after each meeting.

Action items flow directly into the Task Manager app, assigned to the right people with due dates pulled from the conversation — no second copy-paste step, no items lost in a chat thread.

Every meeting is archived in a searchable history. When someone asks what was decided last month, you search for it and find the exact moment — instead of hunting through Google Docs or Slack messages.

Starch connects directly to Google Calendar via scheduled sync, so it knows your upcoming meetings, can surface notes from the last meeting with the same person before you join, and keeps context across calls automatically.

Connect Notion from Starch's integration catalog and meeting summaries can be written directly to your team wiki — the Knowledge Management app auto-categorizes new content and detects when documentation goes stale.

Describe any customization in plain English — 'after every customer call, post a summary to the #customer-success Slack channel and create a follow-up task for the account owner' — and an agent builds that automation without code.

Get closed-beta access →

Toolkit