Can I run this on a live transcript during the meeting?

Better at the end. Mid-meeting the decisions are still moving, and a decision captured too early is the one people later dispute.

What if the notes are my own rough scribbles?

It works, with one caveat: your scribbles already contain your interpretation. The quote rule is weaker here, so mark decisions you cannot quote as OPEN and confirm them with someone.

Workflow automation

Meeting notes to owners

Turn a transcript or scribbled notes into four fixed sections, where an unassigned action shows up as unassigned instead of being handed to whoever spoke last.

Use this within an hour of any meeting where something was supposed to be decided.

The pattern

Copy this and replace everything in [SQUARE BRACKETS] with your own specifics.

Prompt template

Below are notes from a [WHAT KIND OF] meeting on [DATE], between [WHO, with roles].

[PASTE THE TRANSCRIPT OR NOTES.]

Produce exactly these four sections, in this order:

DECISIONS — only what someone explicitly agreed to. One line each. Quote the words that settled it.
ACTIONS — person | what they will do | by when. Write "UNASSIGNED" if no one took it, and "NO DATE" if no date was given. Never guess either.
OPEN — questions raised and not answered, each with who can answer it.
PARKED — raised, explicitly deferred, worth revisiting.

Rules:
- Anything discussed but not agreed goes in OPEN or PARKED, never in DECISIONS.
- Do not merge two people's suggestions into one action.
- Do not add an action that nobody said out loud, even if it obviously follows.
- Nothing outside these four sections. No summary paragraph.

Why each part is there

The four fixed sections: The same four every time, in the same order, which is what makes these notes skimmable in ten seconds and comparable across weeks.
UNASSIGNED and NO DATE: The two most useful strings in the whole pattern. An unowned action displayed as unowned gets picked up; an unowned action quietly assigned to someone gets silently dropped.
Quote the words that settled it: Makes each decision auditable, and makes it obvious when the meeting only felt like it decided something.
Do not add an obvious action: Models fill gaps helpfully. An action nobody spoke is an action nobody committed to, however sensible it looks on the page.
No summary paragraph: The summary is where the careful distinctions above get flattened back into confident prose.

Before and after

The lazy version

Summarise these meeting notes and list the action items.

What goes wrong: Every action gets an owner, because a list of actions without owners looks unfinished. Two of those owners never agreed to anything, and the item nobody wanted has been assigned to the person who happened to mention it.

The pattern, filled in

Worked example

Below are notes from a project sync on 14 March, between me (student lead), two teammates and our mentor.

[transcript pasted]

Produce exactly these four sections, in this order:

DECISIONS — only what someone explicitly agreed to. One line each. Quote the words that settled it.
ACTIONS — person | what they will do | by when. "UNASSIGNED" if no one took it, "NO DATE" if none given.
OPEN — questions raised and not answered, each with who can answer it.
PARKED — raised, explicitly deferred, worth revisiting.

Rules:
- Discussed but not agreed goes in OPEN or PARKED, never DECISIONS.
- Do not merge two people's suggestions into one action.
- Do not add an action nobody said out loud.
- Nothing outside these four sections.

What that buys you: Two real decisions with quotes, four actions of which one is honestly UNASSIGNED — the dataset labelling nobody volunteered for — and the blocker sitting in OPEN with the mentor's name against it. You can send this in one minute and it is checkable.

Now do it on your own work

Run this on your next real meeting, then send the output to someone who was there and ask them to mark anything they disagree with. Count the corrections.

Check yourself — every box should be true

Every DECISION carries a quote from the notes.
At least one item is honestly UNASSIGNED or NO DATE — or you can confirm every action genuinely had both.
You sent it to another attendee and collected their corrections.
You can name the one place the model over-reached, and the rule that should have stopped it.

Then explain it back

Displaying an action as UNASSIGNED felt worse than assigning it to somebody. Why is the honest version more useful to the group a week later?

Write two or three sentences in your own words. If you cannot, you copied a prompt rather than learning a pattern — which is the whole difference this library is trying to make.

Does this depend on which AI tool I use?

The tool choice that matters here is the transcript, not the assistant: automatic transcripts routinely mis-attribute speakers, and this pattern trusts attribution completely. Skim the speaker labels around anything important before you rely on an owner. Also strip names and anything personal before pasting into a hosted tool if the meeting was sensitive.

Why it actually works

Reliable output from a fixed input shape, with the uncertain cases surfaced rather than filled in, is the core design problem in agentic AI — where a wrong assumption becomes a wrong action.

Where agents get this wrong →

Prompting well is a skill you can pick up in a week. Understanding what is happening underneath is the 18-month curriculum — and it is what stays useful when the tools change.

FAQ

Can I run this on a live transcript during the meeting?: Better at the end. Mid-meeting the decisions are still moving, and a decision captured too early is the one people later dispute.
What if the notes are my own rough scribbles?: It works, with one caveat: your scribbles already contain your interpretation. The quote rule is weaker here, so mark decisions you cannot quote as OPEN and confirm them with someone.

Next patterns

The Prompt Patterns Library is a separate, optional track — it does not change the 78-week curriculum or your roadmap. See what the curriculum actually covers →