Claude Prompts for Meeting Notes and Action Items

The short version: Most Claude prompts for meeting notes stop at summarising what was said, which is the easy 20% of the job. The useful 80% is forcing Claude to name an owner and a date for every action item, flag the vague ones that never got assigned in the room, and separate decisions from discussion. Below are the prompts that do that, plus where this still breaks.

Why most meeting note prompts produce useless output

I've watched a lot of people paste a transcript into Claude and type "summarise this meeting" and then wonder why the output is a wall of paragraphs nobody reads. The problem isn't Claude. It's the prompt. A summary of what happened in a meeting is not the same thing as a record of what needs to happen next, and if you don't ask for the second thing specifically, you won't get it.

Here's the uncomfortable bit nobody wants to say out loud: a lot of meetings don't produce clear action items in the first place. People say "I'll look into that" or "let's circle back" and everyone nods, and no AI tool on earth can turn vague human avoidance into a crisp owner and deadline that didn't exist. The best Claude can do is flag that the action item is mushy and make you go back and pin it down with the person who said it. That's valuable, but it's not the "set and forget" magic a lot of blog posts promise.

The core prompt structure that works

Every decent meeting notes prompt has four parts, in this order:

• Context: what kind of meeting, who's in it, what decisions needed to be made
• Format instruction: exactly what sections you want back, in what order
• Strictness rule: tell Claude what to do when information is missing or unclear
• Output constraint: length, tone, who's going to read this

Skip any one of those four and the output drifts. I've seen people skip the strictness rule most often, which is exactly why they end up with action items like "discuss further" with no name attached.

The base prompt

Here's a prompt you can copy and adapt:

• "Here is a transcript of a [type] meeting between [roles, not names if you want to anonymise]. Produce three sections: Decisions Made, Action Items, and Open Questions. For every action item, state the owner, the deadline if one was mentioned, and mark it 'UNCLEAR OWNER' or 'NO DEADLINE' if either is missing. Do not invent owners or dates. Keep each item to one line. Write for someone who was not in the meeting."

That last instruction matters more than people think. If you don't tell Claude who's reading it, it tends to write for the people who were already in the room, which means it skips context that outsiders need.

A worked example

Say you run a 12-person marketing agency and you have a Monday pipeline review with four account managers. A typical transcript runs about 40 minutes and 6,000 words once it's off a transcription tool. Dropped into Claude with the base prompt above, here's roughly what you'd get back:

• Decisions Made: Agreed to pause the Facebook ad spend on the Harrington account until the new creative is approved. Agreed the Q1 retainer renewal pricing will go out at the current rate, not the proposed 8% increase.
• Action Items: Sarah to send revised creative to Harrington by Thursday. James to draft the retainer renewal email, OWNER CONFIRMED, NO DEADLINE STATED. Someone to check with finance on the Q1 budget split, UNCLEAR OWNER.
• Open Questions: Whether the new hire starts on the Harrington account or stays floating across three smaller clients was raised but not resolved.

Notice the two flagged items. That's the prompt doing its job, not failing. Those two things need a two-line follow-up message to the room, something like "James, what's the deadline on the renewal email, and who's picking up the finance check-in?" That follow-up takes 90 seconds and turns a note nobody acts on into one that gets closed out.

Prompts for specific meeting types

Client calls

"Summarise this client call. Separate what the client explicitly asked for from what we offered or suggested. Flag anything the client said that sounded like a complaint or hesitation, even if they moved on quickly. List action items with owner and deadline, and mark anything where we made a verbal promise to the client as HIGH PRIORITY."

That last line matters for client work specifically, because the thing that damages trust fastest isn't missing an internal task, it's forgetting something you told a client you'd do.

Internal strategy or planning sessions

"This was a planning session with multiple options discussed, not all of them decided. Produce a section called Decided, a section called Still Debated with the main arguments on each side in one line, and a section called Parked for Later. Do not merge debated points into decisions."

This stops Claude from flattening a messy, unresolved discussion into something that reads as more settled than it was, which is a common failure mode and one that causes real confusion a week later when two people remember the meeting differently.

Interviews and 1:1s

"This is a 1:1 between a manager and a direct report. Do not produce a formal action item list. Instead, summarise what the person raised as a concern, what support they asked for, and any commitment the manager made. Keep it to under 150 words."

1:1 notes need a lighter touch than project meetings. Over-formatting a personal conversation into a task list makes people feel managed rather than heard, and that's worth avoiding.

A short step-by-step for building your own system

1. Record or transcribe the meeting (most video tools now do this automatically)
2. Paste the raw transcript into Claude, don't bother cleaning it first
3. Use a prompt that names the meeting type, not a generic "summarise this"
4. Require Claude to flag missing owners and dates rather than guess
5. Send the flagged items back to the room within the same working day, while people still remember context
6. Store the final notes somewhere searchable, not just in a chat thread that scrolls away

Step five is the one people skip. A good note that nobody can find three weeks later is worth nothing. If you want a layout that makes this searchable and consistent across your team, this meeting notes template gives you a structure to paste Claude's output straight into.

Where this fits with other Claude work

Meeting notes rarely live in isolation. If you're already using Claude for related document work, it's worth building a small toolkit rather than reinventing a prompt every week. A good starting point is this set of Claude prompts for meeting notes that save time, which covers more variations on the base structure above. If your meetings reference long documents, you'll also want prompts for summarising PDFs so the context Claude uses matches what was discussed in the room.

If you're running the same three or four meeting types every week, it's worth turning your best prompt into something repeatable rather than retyping it each time. Claude's Skills feature lets you save a fixed instruction set as a file Claude reuses automatically, and there are some copy-ready Claude Skills examples worth looking at if you want to go that route. For teams prioritising which meetings are even worth this level of process, the RICE framework for AI prompts is a decent way to decide where the effort pays off and where it doesn't.

One more practical point: you don't need the most expensive model for this. Meeting notes are a high-volume, low-complexity task, which makes them a good fit for a faster, cheaper model rather than running every transcript through the top-tier one. If you're deciding which Claude model to point at this kind of work, this breakdown of how to use Claude Haiku 5.5 is a useful place to check before you default to the most powerful option out of habit.

Where AI-generated action items still go wrong

A few failure patterns show up repeatedly, and they're worth knowing before you trust this blind:

• Overlapping speakers in transcripts get merged into one person's statement, which can misattribute a commitment to the wrong owner
• Sarcasm or jokes get read literally, so "sure, I'll have that done by Christmas" can end up logged as a real deadline
• Long meetings over an hour start losing detail from the middle section if the model is working from a rough transcript with timing gaps
• Action items that were technically said but immediately walked back later in the same meeting sometimes still get listed, because the model doesn't always weight the later correction over the earlier statement

None of these are reasons to skip the tool. They're reasons to read the output once, before you send it to the team, rather than forwarding it unread because it looks tidy.

Frequently asked questions

Can Claude tell the difference between a decision and a suggestion in a meeting?

Only if you ask it to explicitly. Left to a generic summarise prompt, Claude will often blend firm decisions and loose suggestions into the same list. Add a line asking it to separate "Decided" from "Discussed but not agreed" and the output gets much more honest.

Should I let Claude assign action items on its own?

No. Tell it never to invent an owner. If the transcript doesn't make clear who's doing something, you want Claude to flag that gap, not guess based on who talks the most or who said something adjacent to the task.

What's the fastest way to turn a transcript into action items?

Paste the raw transcript, use a prompt that names decisions, action items with owner and deadline, and open questions as three separate sections, and have Claude flag anything missing an owner or date. That one prompt covers most meeting types with small tweaks.

Do I need to clean up the transcript before using it?

No. Claude handles messy transcripts with filler words, interruptions and timestamps fine. Cleaning it first mostly wastes your time, since the model is already good at ignoring the noise and pulling out the substance.