How to Use AI to Turn Meeting Notes Into Something People Actually Read

Last Tuesday I watched my project manager paste raw meeting notes into Claude and get back something that looked like it was written by a corporate communications team having a bad day. Bullet points everywhere. Action items so generic they could apply to literally any meeting in history. The standard AI summary output that everyone ignores.

I've spent the past three weeks testing different approaches to this exact problem. Not because I love meetings — I actively avoid them — but because the gap between "notes exist" and "people actually know what happened" is where most projects quietly die.

The Default Output Is Useless (And Why)

Here's what happens when you paste meeting notes into Claude with a basic "summarize this" prompt: you get a summary. Technically accurate. Completely ignorable. The AI extracts what was said without understanding what mattered.

I tested this with notes from twelve different meetings across three weeks. Client calls. Internal syncs. Planning sessions. Every single default summary had the same problem: it treated all information equally.

The breakthrough came when I stopped asking for summaries entirely. Instead I started asking Claude to identify the one decision that changes what someone does tomorrow. That constraint forced useful output.

My actual prompt structure now: "Read these meeting notes. Tell me the single most important thing someone who missed this meeting needs to know to do their job correctly this week. Then list any deadlines with specific dates. Then list anything that's still unresolved."

The difference is dramatic. Instead of eight bullet points nobody reads, I get a two-sentence opener that actually communicates stakes.

The Format That Gets Opened

I was wrong about this. I thought detailed notes were better. More context. More nuance. Turns out people don't read detailed notes. They skim. They look for their name and any dates.

After testing different output formats on my team — literally tracking who replied to my meeting summaries versus who ignored them — here's what worked:

• A subject line that names the decision, not the meeting ("We're launching Tuesday, not Friday" beats "Marketing sync summary")
• First line answers: why should I care about this?
• Names in bold next to their specific next steps
• Anything unresolved gets its own section called "Still TBD" — this actually generates responses

The "Still TBD" section was accidental genius. People who ignored every other meeting summary suddenly started replying when they saw open questions that affected them. Humans respond to uncertainty more than they respond to decisions.

I now ask Claude to explicitly create this section even when the notes don't obviously contain unresolved items. "Based on these notes, what decisions were implied but not explicitly made? What might cause confusion later?" The AI is surprisingly good at finding these gaps.

The Kick: Voice Memos Beat Typed Notes

This is the thing I couldn't have predicted without testing it extensively.

When I type meeting notes and feed them to Claude, I unconsciously clean them up. I remove the tangents. I organize chronologically. I edit out the weird asides and the moment someone changed their mind mid-sentence.

But those messy parts contain the actual meaning.

I started recording voice memos immediately after meetings — just me talking for 90 seconds about what happened — then transcribing them and feeding that to Claude. The output quality jumped noticeably.

Why? Because in my voice memos I say things like "Sarah seemed hesitant about the timeline even though she agreed" and "I think the real issue is the budget but nobody wanted to say it." Those observations aren't in typed notes. They're the context that makes summaries useful.

Claude can't read body language. But it can work with your interpretation of body language if you actually include it. The AI is only as good as the subtext you give it.

My exact post-meeting routine now: end call, immediately record voice memo while walking to get coffee, transcribe using the phone's built-in transcription, paste into Claude with my prompt template. Takes four minutes total. The summaries are genuinely useful now.

What Still Doesn't Work

Real-time transcription tools that claim to auto-summarize meetings? I've tested three of them. They capture words accurately but miss meaning consistently. The summary of a meeting where someone quit is technically the same as the summary of a meeting where someone got promoted — just a list of topics discussed.

Long meetings also break this system. Anything over 90 minutes produces summaries that are either too long to read or too compressed to be useful. I haven't solved this. My current workaround is treating each major topic as a separate "meeting" for summarization purposes, but that adds friction.

And there's a subtle failure mode I keep hitting: Claude sometimes invents clarity that didn't exist. If the meeting was genuinely confusing and ended without resolution, the AI will still produce a clean summary that implies decisions were made. I've started adding "flag any moments where the notes suggest confusion or disagreement" to catch this, but it doesn't always work.

The unsettling part is how good these summaries look even when they're wrong. Someone who wasn't in the meeting would never know. Which makes me wonder how many AI-generated meeting summaries are currently circulating in organizations, looking authoritative, containing subtle misreadings that nobody catches because nobody compares them to what actually happened.

I still use Claude for every meeting summary. But I've stopped trusting it completely, which feels like the right relationship to have with any tool that's better at sounding confident than being correct.

Heads up: Some links in this post may be affiliate links. I only recommend tools I've personally tested. Opinions are entirely my own.