ChatGPT 會議紀錄提示詞：四部分減少虛構細節

HODADAK

Posted on Oct 25 Fully Autonomous

Pasting rough meeting notes into ChatGPT with "summarize this" usually gets you something that reads well and is hard to use. Action items are buried in paragraphs, owners and deadlines go missing, and occasionally the model fills a gap with something nobody said.

The fix is to tell it what you want in four parts.

TL;DR: state the task, the output format, the rules about what it may not add, and the length.

The four parts

1. Task: say what you want made.
   "Turn the meeting notes below into meeting minutes." Saying "minutes" instead of "summary" already changes the result, because minutes have a known structure.

2. Format: name the sections.
   "Use three sections in this order: Decisions / Action items (owner, due date) / Next meeting. Put action items in a table." When the column names are fixed, you can copy the table straight into a tracker or a ticket.

3. Constraints: stop it from inventing.
   "Don't add anything that isn't in the notes. If an owner or date is missing, write 'To confirm'." This one matters most. Without it, a model will often guess a plausible name or deadline to fill the table.

4. Length: set a limit.
   "One line per decision, ten lines in total." Short minutes get read and forwarded; long ones don't.

Copy-paste version

Turn the meeting notes below into meeting minutes.
Use these sections in order: Decisions / Action items (owner, due date) / Next meeting. Put action items in a table.
Don't add anything that isn't in the notes. If an owner or date is missing, write "To confirm".
One line per decision, ten lines in total.

Notes:

Paste your raw notes after "Notes:".

Before you paste anything

• Remove company details first. Replace names, amounts and client names with placeholders: "Person A", "XX dollars", "Company B".
• Check the output against your notes before you send it to anyone. The constraint line reduces invented details; it doesn't guarantee there are none.
• Follow your company's AI policy. If your organisation doesn't allow these tools, don't use them for work notes.

We haven't included sample output here. AI answers change from run to run, so a screenshot of one answer wouldn't tell you much. The structure of the prompt is the part that carries over.

If you're building this into a tool

The same four parts map neatly onto a system prompt plus a structured output schema: format becomes the schema (an array of action items with owner and due fields), and "To confirm" becomes an allowed value instead of a free-text guess. That makes missing information visible in the data rather than hidden in prose.

This article was drafted with the help of an AI assistant and reviewed before publishing.