Turning Messy Notes into Actions — AI for Work & Productivity

You'll learn

The difference between summarizing and extracting
The four things worth pulling from any notes: decisions, owners, next steps, open questions
How to build a personal extraction prompt you reuse everywhere
How to quality-check what the AI pulled out

01Section

Extraction is not summarizing

In Summarizing Emails and Notes, you compressed — long thing in, short thing out. Extraction is a different job: pulling specific, structured items out of the mess and throwing everything else away. A summary answers "what happened?" Extraction answers "what happens next, and who's on it?"

The source can be anything rough: scribbled meeting notes, a call transcript, a rambling voice memo, the whiteboard photo you typed up. The messier the input, the more this skill pays.

Plain-language definition

Extraction — pulling specific, structured items — decisions, tasks, names, dates — out of unstructured text and discarding the rest. Done right, every item in the output traces back to a line in your source.

Checkpoint

02Section

The four things worth extracting

Whatever the notes are from, the same four things are buried in them:

Decisions
What got settled. These stop being discussion topics and start being facts.
Owners
Who is responsible for each item. One name per action — a task owned by "the team" is owned by no one.
Next steps
The concrete actions, each with a deadline if one was said out loud.
Open questions
What was raised but never resolved. These become the next conversation's agenda.

Quick knowledge check

Your notes say: "Budget approved. Someone should tell finance. Timeline TBD." What's still open here?

AThe budget approvalBThe timeline — and "tell finance" has no owner yetCNothing — all three items are resolvedDThe notes are too messy to use

Checkpoint

03Section

Build your personal extraction prompt

A generic extraction prompt works. A personal one works better, because your notes have patterns: the same recurring meetings, the format your task manager wants, the category you always need (like "things I promised people"). Build it once, refine it on real notes, and keep it somewhere you can paste from in five seconds.

Starter template — make it yours

Extract from my notes below, exactly as stated — don't infer or invent:
1. DECISIONS — things that were settled
2. ACTIONS — each as "[owner]: [task] — [deadline]". If no owner or deadline was stated, write "UNASSIGNED" or "no date" instead of guessing.
3. OPEN QUESTIONS — raised but not resolved
If a category is empty, say "none." Quote the original wording where it matters.

My notes:
[PASTE NOTES HERE]

Run it on real notes: Use notes from this week, not a made-up example — real mess is the test.
Notice what's off: Wrong format? Missing a category you always need? Too wordy?
Edit the prompt, not the output: Fix it once at the source so every future run comes out right.
Save it somewhere fast: Notes app, text shortcut, saved prompt — five seconds from messy notes to action list.

Privacy note: Rough notes are the least filtered thing you write — full of names, half-decisions, and candid asides. Do a ten-second sweep before pasting: initials instead of names, and cut anything you wouldn't want forwarded.

Checkpoint

04Section

Check what it pulled out

Extraction feels mechanical, and that's the trap: the output looks so clean and confident that you forget to check it. Three failure modes show up regularly:

It misses items. A decision phrased casually — "let's just go with the blue one" — can slip straight past it.
It invents owners. If Sam was mentioned anywhere near a task, the AI may confidently assign the task to Sam. Nobody did.
It upgrades maybes into facts. "We could aim for March" can come out as "Deadline: March."

The invented owner, in the wild

Notes: "Discussed the onboarding doc. Sam mentioned the old version is outdated. Need to rewrite it." AI output: "Sam: rewrite the onboarding doc." Sam pointed at the problem — nobody assigned the fix. If this lands in the follow-up email, Sam learns about their new project by surprise.

The check is fast because you own the source: read each extracted action and find the line in your notes that backs it. A minute, maybe less — and it's the difference between a tool you trust and one that quietly hands work to the wrong people.

Key takeaway

Every extracted item should trace to a line in your notes. If you can't find the line, the AI wrote it.

Pause and think: Think of the last meeting notes you took. If AI had extracted the action items, would you have checked them — or forwarded the clean-looking output as-is?

Checkpoint

Prompt exercise

Extract, then grade it

Find real messy notes — from a meeting, a call, or a rambling voice-memo transcript — and sweep out sensitive names. Copy this prompt into ChatGPT, Claude, Gemini, Copilot, or whichever AI tool you have access to (the website doesn't run AI itself), paste your notes in, then grade the output against the source like a teacher marking homework.

Extract from my notes below, exactly as stated — don't infer or invent:
1. DECISIONS — things that were settled
2. ACTIONS — each as "[owner]: [task] — [deadline]". If no owner or deadline was stated, write "UNASSIGNED" or "no date" instead of guessing.
3. OPEN QUESTIONS — raised but not resolved
If a category is empty, say "none." Quote the original wording where it matters.

My notes:
[PASTE NOTES HERE]

Reflection: Score it honestly: anything missed? Any owner or deadline that isn't actually in your notes? Whatever it got wrong is exactly what your personal version of this prompt should guard against — go edit the prompt now, while the miss is fresh.

Quick check

3 quick questions — no pressure

There's no pass or fail here. Answer them all, and we'll show you the answers either way.