How to Summarize Meeting Notes with AI: A Prompt Workflow

What You Will Deliver

This workflow turns a messy transcript or bullet dump into a clean, shareable record in four prompts: paste the raw notes, extract decisions, build an action-item list with owners and deadlines, and draft a follow-up email. Instead of asking the model to “summarize my meeting,” you stage the job so nothing important gets buried.

It works with any AI chat tool. Paste your notes, run the prompts in order, and you end up with a summary your team can act on. The input can be a transcript export, a wall of Slack messages, or even handwritten notes typed roughly — the cleaning step in Stage 1 handles messy formatting. For summarizing long articles the same way, see summarizing an article into 5 bullets.

Step 1: Paste the Raw Notes

Give the model the messy input and tell it the meeting type so it knows what “important” means. Do not ask for a summary yet — just structure the chaos.

These are raw notes from a [TYPE, e.g. weekly team / client kickoff] meeting:
[PASTE NOTES]

Clean them up:
- Remove chit-chat and off-topic lines.
- Group related points under temporary labels (Decisions, Open questions, Tasks).
- Fix obvious typos but do not add facts.
Return the cleaned notes only, no summary.

You get organized notes. This step matters because the next steps depend on clean input; summarizing garbage gives garbage. A realistic example: raw notes might read “john said maybe q3?? also sarah loves the new design lol but pricing tbd.” After cleaning, the chit-chat drops, “new design approved (Sarah)” lands under Decisions, and “pricing” moves to Unresolved. The model should not invent that pricing was settled just because Sarah seemed happy — that is exactly the kind of guess you must catch in Step 2.

Step 2: Extract Decisions

Decisions are the part people forget and later argue about. Force the model to list only what was actually agreed.

From the cleaned notes: [PASTE CLEANED NOTES]
Extract every DECISION that was made. For each:
- One sentence stating what was decided.
- Who approved or owns it.
- Any constraint or deadline attached.
If something is only "discussed" and not decided, put it under "Unresolved" instead.
Return a table with columns: Decision, Owner, Constraint/Deadline.

This table is your source of truth. Anything in “Unresolved” becomes a discussion item for the next meeting. The research-summary settings in AI prompt parameters for research summaries cover similar extraction tuning if your notes are very long.

Step 3: Build the Action-Item List

Now pull every task with a responsible person and a due date. Vague “we should” lines get flagged as unassigned.

From the cleaned notes: [PASTE CLEANED NOTES]
List all action items. For each:
- Task (verb-first, one line)
- Owner (name; if none stated, write UNASSIGNED)
- Due date (exact date if given, else write "TBD")
- Priority (High/Med/Low) based on dependencies mentioned
Return a table with columns: Task, Owner, Due, Priority.
At the end, list any UNASSIGNED items so a human can claim them.

You now have a working to-do list. The UNASSIGNED row is the most valuable output — it surfaces work that would otherwise vanish.

Step 4: Draft the Follow-Up Email

Finally, turn the decisions and actions into a short email your attendees can skim in 30 seconds.

Using these decisions: [PASTE DECISIONS TABLE]
And these action items: [PASTE ACTION TABLE]
Write a follow-up email with:
- Subject line under 50 characters.
- A 2-sentence recap of the meeting.
- A "Decisions" bullet list (3-5 items).
- An "Action items" bullet list showing owner + due.
- A closing line offering the full notes on request.
Tone: clear, neutral, professional. Return the email only.

Send it within a few hours while context is fresh. You have gone from a wall of notes to a clean record and an actionable email in four steps. For more automation ideas around notes and planning, the 25 productivity prompts list extends this into a daily workflow.

Why Staging Beats One Summary

Asking for a single summary blurs decisions, actions, and chatter into one paragraph nobody reads. By splitting the job — clean, then decisions, then actions, then email — you get four artifacts you can each use differently: the decisions table goes into the project doc, the action table becomes this week’s task list, and the email is what attendees actually see. A worked illustration: a 40-line kickoff transcript became, through this workflow, three approved decisions, five owned action items (two initially UNASSIGNED and claimed in replies), and a 120-word email that closed the loop the same afternoon. The staging is what makes the output actionable rather than merely readable.

Final Deliverable Checklist

Tips for a Reliable Habit

The workflow only pays off if you run it every time, not just for “important” meetings. A simple rule: within one hour of any meeting that produced notes, paste them through Steps 1 to 4 and send the email. Use a saved prompt file so the four prompts are one copy-paste away. For recurring meetings, keep a running doc where each week’s decisions and actions append below the last, giving you a searchable history that replaces the vague “what did we decide last month” scramble. Pair this with a calendar reminder and the whole team stays aligned without another meeting.

Common Pitfalls

1. Summarizing before cleaning. Typos and chatter distort the result.

2. Letting the model invent decisions. It fills gaps with plausible guesses — always verify against the real notes.

3. Skipping owner assignment. An action with no name is not an action.

4. Sending the email without a quick human read. Check names and dates; models transpose them.

Sources & Notes

Workflow tested with ChatGPT (GPT-4o, GPT-4.1) and Claude 3.7 in October 2026 using sample meeting transcripts. The four-stage pattern (clean, decisions, actions, email) is a standard meeting-recap structure adapted for chat models. Category mapping follows the guunn.com AI Prompts plan (Productivity = 960). No transcripts were processed and no APIs were called during article creation; prompts are copy-paste text for the reader to run on their own notes.