Extract Action Items From Documents in ChatGPT

Meeting notes and briefs are full of implied work. Someone will follow up on the vendor. Someone should draft the announcement. Someone owes a decision by Friday. Reading past those lines is easy. Turning them into a real list of who does what by when is where the value shows up.

ChatGPT can do this well when you ask for a specific shape. You want each action with an owner and a deadline where possible, plus a separate list of open questions. Left to itself, ChatGPT tends to blur these together and skip the owners. Ask for them by name.

What counts as an action item

An action item is one clear task, an owner, and a deadline or trigger. If any of the three is missing, it is not really an action yet. Sometimes the source is missing the owner or the deadline, and that is fine. Marking the gap is more useful than inventing an owner who never agreed.

When this technique is useful

• Weekly meeting notes that mix decisions and to do items
• A project brief that hides tasks inside descriptive paragraphs
• Interview notes from customer research with follow up ideas
• A postmortem report where fixes need owners and dates
• An email thread you exported as a single document

Prompt for weekly meeting notes

File: weekly_meeting_notes.md.
Task: extract action items and open questions.
Return two lists.
List 1 action items: task, owner, deadline, source line.
List 2 open questions: question, who raised it, source line.
If owner or deadline is missing, write "not stated". Do not invent either.

Why this prompt works

The two list split keeps decisions separate from questions. The source line column makes each item traceable back to the notes. The not stated rule protects you from fake owners, which is the single most common failure of this task.

Prompt for a project brief

File: project_brief.pdf.
Task: find every task described or implied in the brief.
Return a table: task, owner if named, deadline if named, section heading, quoted supporting text.
Sort by section order in the brief.
Skip tasks that are already marked as done in the source.

Prompt for a postmortem report

File: incident_postmortem.md.
Task: pull out every follow up action.
Return a table: action, category from prevention, detection, or response, owner if named, target date, why it matters.
If a category is unclear, mark it "unclear".
Group identical actions under one row and list all owners in a bullet inside the cell.

What to include in every request

• A clear ask for actions and questions as separate lists
• Column names, including a source line or section reference
• Explicit rules for missing owner or deadline
• A rule against inventing owners or dates
• A note about which items to skip, such as already done tasks

How to refine the list

Read the first three rows and check them against the source. If the source line is missing or wrong for one row, ask ChatGPT to redo that row with the exact quoted sentence and section. Row by row refinement is faster and safer than restarting the whole extraction.

Common mistakes

• Accepting a task with no owner listed as if the owner is you
• Trusting invented deadlines that sound reasonable
• Mixing decisions, actions, and questions into a single list
• Skipping the source line column so you cannot verify anything
• Rebuilding the list every time instead of correcting rows

How to check the result before you share it

Read the list as if you were the person named on each row. Would you know exactly what to do next. Would you know when it is due. Would you know where the task came from. If any answer is no, ask for more detail on that specific row.

Takeaway

Actions with owners and deadlines get done. Actions without them get lost. Ask for both by name and mark the gaps honestly.