Start here Start here 4 min read
A familiar scene

'Summarize this' gives you a neutral digest that misses what mattered to you; naming what you're looking for makes it hunt through fourteen pages for exactly that

Don't ask AI to 'summarize the meeting' — tell it to find the things you actually need

Cluster11 · Start here
Topicscontext · instructing · beginners
TL;DR

"Summarize this" is an open-ended instruction: you never said what you care about, so the AI returns a neutral digest — correct, complete, and likely to skip the one piece that mattered to you. Change the ask from "summarize" to "find me these four things, and quote the exact line," and it flips: the machine hunts fourteen pages for what you need, while the judgment of "what to do next" stays yours.

The two-hour meeting just broke up. People are standing, closing laptops, agreeing on where to get lunch. You're still sitting there with a fourteen-page transcript, needing to pull five sentences out of it for the people who missed it.

You paste the whole thing into AI and type "summarize this." Ten seconds later it hands you something tidy and correct — and you still feel short-changed. The decision the boss locked in around minute eighty, the task that quietly landed on you near the end: it either folded them into one bland line or skipped them. So you open the transcript again, to re-read the thing you just sat through for two hours.

01"Summarize" makes it guess what you care about — and it guesses evenly

It isn't that the AI failed to read it. It swallowed those fourteen pages in one gulp. The break is that you never told it what you needed.

"Summarize this" is like handing the transcript to someone who wasn't in the room and saying "boil it down." All they can do is guess what's worth keeping — and the safest guess is to flatten everything evenly: one line per topic, nothing emphasized, nothing dropped. The result is correct for everyone and sharp for no one — least of all you, who walked in with a very specific concern you never said out loud. The flatness isn't laziness. It's exactly what you ordered when you left "for whom, and to do what" blank.

02Don't ask it to summarize — ask it to find

Here's the turn: you don't actually want a summary. You want to know four or five specific things. So ask for those four or five things directly, instead of routing the question through the word "summarize."

14-PAGE TRANSCRIPT5 LINES WORTH SENDING
Stage 1 — You name what you need: "Read this transcript and find me four things — what got decided, who owns what, what's still unresolved, and anything that touches my part." You set the target instead of letting it guess.

Stage 2 — The machine hunts and quotes: it scans all fourteen pages, gathers exactly those four buckets, and quotes the original line for each — so you can trust it without re-reading from the top.

Stage 3 — You decide: looking at the filtered table, you choose which lines deserve a spot in the five you send. The judgment is still yours — it's just done on one clean page instead of fourteen raw ones.

The difference isn't the AI "trying harder." It's that you changed the question from "boil it all down" — which it has to guess at — to "find me these specific things" — which is what it's best at.

Notice the word quote in stage two. This is where a long-but-useless summary splits from a usable one. When you make it cite the original wording behind each point, you get two things at once: you can check fast whether it made anything up, and you keep the original nuance instead of a paraphrase that's been sanded down. A summary with no sources is one you still have to trust on faith.

03Four framing questions, paste and go

You don't need any "technique." You just say out loud the four things you were already scanning for every time you read minutes:

1
Decisions

"List every decision that was locked in — and quote the exact line where it was made."

2
Tasks and owners

"Who took what, by when? Only pull the ones with a clear owner — don't infer."

3
Still open

"What was raised but not settled? Just list it — don't answer it for me."

4
Touches my part

"Anything that mentions my piece of the work? Pull it, with a few lines around it."

These four aren't prompt tricks. They're the four questions you already ask yourself when reading minutes — now said out loud, for a machine that can't read the thoughts in your head.

The third one is the one to watch: "just list it — don't answer it." If you let AI settle a still-open question on its own, you've handed it the thing it's most likely to get wrong — judgment based on context it doesn't have — and brought home an answer that sounds reasonable but that nobody in the room actually agreed to. Making it stop at "here's what's unresolved" keeps it on the right side: the machine clears the table for judgment, it doesn't sit down in the judgment seat.

So next time you're staring at a long transcript, don't open with "summarize this" and then feel let down by the flatness. Ask it straight for the things you actually need to find — that's when a forty-minute job shrinks to a few minutes. It's also where you brush up against a bigger skill: choosing what to put in and what to have the agent pull out, and giving a structured instruction instead of an open-ended one. That fourteen-page transcript never needed you to read all fourteen pages. It only needed you to say, once and clearly, what you came looking for.

Other pains

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.