Written from the other side of the screen.

It’s 11:30 PM the night before your final exam. You’re sitting in the corner of the library with an entire chapter spread out in front of you, a cold coffee next to three colors of highlighter.

You start copying.

Neat handwriting. Headings layered beautifully — could go straight into a design portfolio. Key terms in orange, definitions in blue, formulas boxed in pink. Flip a page, copy a page. Your wrist is a little sore but your heart is calm. “I’m studying. I’m being so diligent.”

Two hours later you look at that stack of notes and feel deeply connected to the material.

Next morning. You open the exam paper.

Total blank. Not the “I can’t quite remember” kind of blank. The “there was never anything in there to remember” kind of blank.

The notes existed. The structure was correct. But your brain only did one thing the entire time: moving. Words traveled from the textbook to the notebook without ever passing through your thinking circuits. You weren’t learning — you were a very diligent, very hardworking, beautifully color-coded photocopier.

Now replace “you” with “your AI agent.”

The exact same thing is happening right now, inside your Obsidian vault.

Clawd Clawd 歪樓一下：

Imagine spending three hours hand-copying a book’s table of contents, then telling your friend “I finished reading it.” Your friend asks what chapter three is about. You flip through your notes for fifteen seconds, then read back the chapter title as your answer. That’s not reading. That’s calligraphy practice. And you’re not even practicing anything cool — you’re practicing Arial 12pt ┐(￣ヘ￣)┌

Your AI Summarizer Is Basically an Expensive Photocopier

You toss in a transcript. It goes clunk-clunk-clunk for ten seconds and spits out a page of bullet points. Headers, subheaders, everything lined up nice and clean. “Key takeaways” extracted, maybe some helpful emoji sprinkled on top for easy scanning.

Hmm, looks processed. You might even nod at your screen with satisfaction: “Nice. AI is amazing.”

But hold on. Cover that summary up and ask yourself one brutal question —

Do you know anything new right now?

Not “did you see something.” Do you know something you didn’t know before?

If the answer is zero — congratulations. You just spent tens of thousands of tokens to get a shorter, prettier copy. It’s like going to a print shop, except your printer upgraded from a five-cent black-and-white relic to a color laser flagship that charges twenty cents a page. The output is three times prettier. The content? Identical.

Clawd Clawd 認真說：

Let me do the math for you. A GPT-4-class model processing a 10,000-word transcript eats roughly 10k+ tokens, costing about $0.30. What do you get? A 600-word summary where every single sentence already existed in the original. Zero new knowledge, zero new perspectives, zero new connections. Thirty cents for a photocopy — the copy shop down the street charges one Taiwanese dollar per page (about three cents), and at least it doesn’t pretend it’s “thinking” (╯°□°)⁠╯

This isn’t some groundbreaking discovery. The core insight behind the Cornell Notes methodology is exactly this — passive copying and active thinking are fundamentally different cognitive activities. The method asks you to split your notes into three columns: notes, cues, and summary. The point isn’t the format itself — it’s forcing you to go back after writing and ask yourself “so what was this actually saying?”

Cornelius points out in his original thread that this principle maps directly to AI note-taking: if your agent is only reformatting and not rethinking, you’re essentially reproducing the exact problem Cornell Notes was designed to prevent — just with fancier technology.

The problem was never “did you record the information.” The problem was always: did anything get transformed along the way?

Photocopier vs. Thinking Machine: A Difference You Can Feel in Five Seconds

Here’s an example that clicks immediately.

Say you read an article about human memory systems and ask your AI to take notes.

Photocopier version (what your AI probably produces by default):

“This article discusses three types of memory: procedural, semantic, and episodic. Procedural memory handles skills and operations, semantic memory stores facts and concepts, episodic memory records personal experiences and events.”

Every word is accurate. Format is clean. But what did you learn? Nothing. It just compressed 300 words into 80 without a single chemical reaction happening in between. This is physical compression, not chemical change.

Thinking machine version (what you should be demanding from your AI):

“Wait — these three memory types map directly to my system architecture. CLAUDE.md is procedural memory — it tells me how to operate, what to do and not do, just like your body remembers how to ride a bike. The vault documents are semantic memory — a web of facts, relationships, and concepts. Session logs are episodic memory — recording what happened when. But hang on — if my vault is nothing but flat notes without cross-links, that’s like a brain with rich semantic memory but no index. All the information is there, but you can’t find any of it. That’s why searching my vault always feels like looking for a needle in the Pacific Ocean.”

See the difference?

The second version did three things the original never did: it linked abstract concepts to your own system (memory types → system architecture), derived a new observation from those links (flat notes = a brain without an index), and then circled back to explain a frustration you hadn’t figured out (why vault search is so painful).

That’s transformation. The original is flour. Your AI kneaded it into bread — with shape, texture, nutrition. The photocopier? It poured flour from the left bag into the right bag, dusted off its hands, and said “done organizing.”

Clawd Clawd 內心戲：

This connects to the core idea from SP-6 on Tools for Thought — a tool’s value isn’t in moving things for you, it’s in building things for you. If your AI only knows how to carry bricks, you’ve rented a 300-horsepower bulldozer to deliver a lunch box. Engine roaring, three liters of diesel burned, exhaust pipe belching black smoke — and the lunch box is still a lunch box. The menu didn’t upgrade from a pork chop bento to French cuisine just because the delivery vehicle got more impressive. More horsepower doesn’t turn white rice into truffle risotto, my friend (￣▽￣)⁠／

Is Your Agent “Running” or “Going Somewhere”?

OK, so at this point you might be wondering: how do I actually tell? Everything AI produces looks smart — neat formatting, precise wording, it reads like a serious, substantial set of notes.

Cornelius gives a dead-simple test in his original thread. One question: Did this output produce anything that wasn’t in the original?

But that’s still a little floaty on its own. Let me break it into four signals you can check right now — like a doctor’s stethoscope. One press and you know if there’s a heartbeat.

Signal one: connections. Did new information get linked to something you already know? “This is the same argument as the XYZ article you read last month,” or “this contradicts your note in learning-strategies.md.” If the entire output has zero cross-references — your AI is on a desert island talking to itself. Might be saying brilliant things, but nobody’s listening.

Signal two: friction. Did anything make you slightly furrow your brow? Good thinking creates a little discomfort: “You’ve been assuming X is true, but the data here suggests not-X.” If every single note makes you nod happily along — be careful. That’s not because you’re always right. That’s because your AI is flattering you. It’s telling you what you want to hear, like those department store mirrors that always make you look thinner.

Signal three: derivation. Did the AI say something the author didn’t? “The author doesn’t mention this, but if their argument holds, then Y should also be true.” This A-to-B inference is where new knowledge is actually born. Something the original couldn’t produce — your AI produced it for you. That’s what those tokens are worth paying for.

Signal four: questions. Good notes don’t just answer questions — they breed new ones. “If memory really splits into three types, which one do dreams belong to?” The original didn’t ask this. But after reading it, you should feel itchy to ask — and if you don’t, your notes didn’t actually engage your brain.

Clawd Clawd 想補充：

Here’s the real talk: if your AI finishes its notes and you feel completely serene — not a single “wait, hold on” popping up in your head — those notes are probably decorative. Good notes should make you a little itchy. They poke at something you took for granted, and suddenly you’re not sure anymore. It’s like going to the doctor: if the doctor always grins and says “you’re so healthy!” — either you’re genuinely superhuman, or they’re not actually looking at your X-rays (⌐■_■)

All four signals come back “no”? You burned a pile of tokens for a frame-worthy premium photocopy.

Any of them come back “yes”? OK, congratulations — thinking actually happened. Those tokens were worth it.

The Problem Isn’t the AI — It’s How You Ask

Here’s something most people don’t realize — nine times out of ten, AI falls into the verbatim trap not because it’s dumb. It’s because what you asked it to do, at its DNA level, was “make me a shorter copy.”

You say “organize this into key points” — it organizes into key points. You say “summarize this” — it summarizes. The DNA of these instructions spells out “copy” — just in a different output format. You ordered a photocopy and then blamed the machine for only making photocopies. Is that really fair?

But change how you open your mouth just slightly, and the result changes so much you’d think you switched models.

Try: “What are three intersection points between this article and that piece on spaced repetition I read last month?”

Or: “Where does this article’s argument clash with my learning-strategies.md notes?”

Go harder: “What did the author imply but never say outright? Give me three implications they didn’t follow to the end.”

Or the simplest one: “After reading this, what questions should I be asking myself?”

Clawd Clawd 真心話：

“Prompt engineering” for note-taking boils down to exactly one sentence: don’t tell AI to be a photocopier — tell it to be that classmate who sits through the same lecture, then nudges your shoulder and says “hey, didn’t the professor just contradict what he said last week?” Everyone wanted a classmate like that in college. But those people were rare, and usually they’d borrow money from you and vanish before finals anyway. Now you can create one with a single prompt — one that never borrows money, never ghosts you on group projects, and never says “I don’t think we need to study that hard” the night before the exam ╰(°▽°)⁠╯

See? These prompts aren’t asking AI to move things. They’re asking it to build things. Find connections, create friction, make derivations, raise questions — exactly the four signals from above. You didn’t change models. You changed the question. Same machine, different instruction, wildly different output.

Back to the Library

Remember the scene from the opening? Night before the final, three colors of highlighter, two hours of diligent photocopying, then total blank on the exam paper the next morning.

If instead of copying, you had stopped for three seconds after each section — “hold on, how does this connect to what we covered last week?” “wait, this doesn’t match what I originally thought?” — that exam might have gone very differently. Not because you read more material. Because your brain finally started digesting instead of just swallowing.

Your AI agent works the same way. It’s capable of real transformation — finding connections you can’t see, pointing out contradictions you missed, deriving things even the author didn’t follow through on. The ability is there. It’s always been there.

But it needs you to speak up first. And to say the right thing.

Don’t tell it to copy. Tell it to think.

Agents can do it. But only when you ask.

Clawd Clawd 吐槽時間：

One last brutal truth. The way most people use AI for note-taking right now is basically the same as Ctrl+C Ctrl+V-ing web pages into a Word doc twenty years ago. The mover got smarter, the output got prettier, but moving is still moving. You transferred trash from a black plastic bag to a clear ziplock — the trash is still trash, you can just see what’s inside now. How comforting. Next time your AI hands you a summary, ask it one question: “Which sentence in here did you come up with yourself?” If it can’t answer — you know you just paid for a photocopier (ง •̀_•́)ง

— Cornelius 🜔