给ChatGPT施咒,让它为你制作Anki卡片— (English)

Generated: 2026-06-23 07:16:26

---

You must have had this experience—full of hope, you dump a pile of material into ChatGPT, and five minutes later, staring at the "masterpiece" on the screen, you feel like yanking it out of the cloud and beating it up.

That was me that time.

I spent a full half hour pasting the original textbook text word by word into ChatGPT, thinking, "Finally, I can let my brain rest and let it do the work." And what did it give me? Dozens of Anki cards. I took one look—man, those answers were written like mini-essays. The front-side questions were so rushed they seemed desperate to get to the afterlife, and the back-side answers were dragging and sloppy, even copying in that useless phrase "It is worth mentioning that..." from the original text, without cutting a single hair.

You're feeling all pleased—"Ah, technology is amazing"—and then reality slaps you awake: This thing isn't here to help you; it's here to waste your life.

I was so furious at that moment, I really wanted to smash the keyboard.

But later I figured it out. Guess what? It wasn't that ChatGPT was stupid—it was that I was naive.

In my understanding, "Help me make a flashcard"—how simple is that? But what does AI think a "flashcard" is? Just one sentence on the front and one on the back, even if the front asks "How many galaxies are in the universe?" and the back answer dares to write three pages. It has no clue what "conciseness" or "efficiency" mean, because you never taught it.

Speaking of which, does this remind you of those fresh high school graduates? You tell them, "Clean up your room," and their idea of "clean" is shoving clothes into the closet and smoothing out the bedsheet, and they're done. What you think "clean" means—that's minimalist decluttering.

So I started a two-week "spell-casting" training. Today I'm spilling the whole process—how you fell into the pit, how I tried and failed, step by step, all written out here.

---

First Version: Just Giving It Requirements—Might as Well Say Nothing

At first, I thought, okay, it doesn't know my standards are high, so I'll tell it:

The answer should be short. Don't repeat the nonsense from the question.

You have to admit, the effect improved a little—at least it stopped copying the question back into the answer.

But look at the cards it made—

Card 3: The concept of "dual task," it stubbornly split into four lines. You don't want to read past the first line. Why? Because it has no idea what "short" really means. In its understanding, "short" might just mean "write three fewer sentences of nonsense," but when it comes to splitting complex concepts? It splits them all the same.

It's like telling a chef, "Make the food delicious," without telling him the spice level, the saltiness, the heat—what does he make? You just have to guess and swallow it yourself.

Second Version: Give Examples—Then It Really Understood

I dug out a perfect example I had written when translating SuperMemo's 20 card-making principles:

Front: What is the smallest unit of memory?

Back: Atomic knowledge point

Just that simple card—even my mom saw it and said, "I get it."

I threw this example into the Prompt and told it, "Follow this style."

Guess what? The effect was like I got a new AI—it started learning to break long sentences into short Q&As, and even added an extra card at key points to cover the topic thoroughly.

But a new problem arose: on the "dual task" point, it still wrote a big chunk of explanation. Why? Because even though it saw my example, it didn't know why I did it that way. The example is just the surface; it never grasped the underlying card-making thinking.

Third Version: Make It Write Out Its Mental Steps

That's when I thought of Chain-of-Thought—CoT for short.

I'd read papers before: if you let the AI write out its thinking process, the results are much better. Since it can reason on its own, why not just make it follow my flow?

So I gave it a standard five-step card-making method:

1. Read through the text and pick out all key terms.
2. Break each concept into the smallest possible knowledge points.
3. For each knowledge point, summarize it in one sentence.
4. Generate a question, ensuring the answer doesn't exceed one line.
5. Output the final result.

This time I left out the example and tried CoT alone. It started consciously summarizing the content during card-making, no longer stuffing explanatory text into the answers.

But a new issue popped up—sometimes it summarized too aggressively, swallowing up key details. Like telling a chef to write a recipe; to simplify, he writes "boil it" instead of "first bring to a boil over high heat, then simmer on low."

Fourth Version: Combine Examples and Steps—The Ace

I put the Few-shot example and Chain-of-Thought steps together. First show it that sample card "Atomic knowledge point," then make it follow the five-step method.

The first set of cards it produced made me slap the table—

• It finally split "as an educational discipline" and "as a psychological discipline" into two separate cards.
• For "dual task," it divided understanding into two levels: one for short-term memory tasks, one for long-term memory tasks.
• Every answer was under 15 characters.

Think about it—15 characters! Better than my handwriting.

But before I could celebrate, another shadow appeared—sometimes in its reasoning steps it wrote the answers too detailed, and then the final card became wordy again.

Later I added one more line: "Keep the thinking process to yourself; the card should only contain the question and answer." That stabilized it.

Fifth Version: Loop to Standardize Format, Don't Let Export Become a Nightmare

Remember that feeling? You make the cards beautifully, ready to import into Anki, and then the format is all messed up.

Sometimes it's a Markdown list, sometimes plain text, sometimes with numbering. And you have to fix it manually—such a headache.

The solution is actually super simple—add one more step in the CoT: "Finally, output in Markdown table format, two columns: Question, Answer, Tags."

Or, replace the example in Few-shot with a table format, and it will follow along.

Also, a little secret: a friend told me to remove all paragraph breaks from the input text, stringing all sentences into one paragraph, and ChatGPT's output would improve a lot. I tried it myself—it really works! Why? I guess paragraphs make the AI pay too much attention to structure, ignoring semantic connections.

Later there was another slick move: delete the last line "A set of cards:" from your Prompt, and ChatGPT will go freeform, producing three or four sets of cards from different angles. But the quality is unstable—sometimes it introduces extra knowledge points. If you want it strictly based on the original text, it's best to give it a clear output framework.

Advanced Tricks—You Can Play with Them Too

Later I looked at tips shared by other people who tinker with this stuff, and learned a few more tricks:

First move: Use GPTs to make a dedicated card-making agent.

Someone directly threw "How to Write Good Cards" and "20 Card-Making Principles" into GPTs' knowledge base, letting the AI read and learn on its own. In practice, the effect was close to my CoT+Few-shot combo, but saved you the time of writing the Prompt.

Second move: Upload files and let it use them as reference.

But careful! If you say, "Make flashcards based on the file I uploaded," it will treat the reference file itself as material. You have to explicitly state in the Prompt: "The following content is a tutorial, not the content to work on."

Third move: Manually process the reference file.

For example, feed it an outline in mind map format—it can generate more reasonable splits based on hierarchy. But you have to be proactive—it's like a lazy chef; if you don't push, it'll only serve you two dishes.

---

Addressing Your Doubts: Why Not Just Write Them Yourself?

Some will say: "You've been messing around for two weeks—it would have been faster to write the cards by hand."

Let me do the math: A 300-page textbook, fully handmade cards, takes about 6 to 8 hours. With my current Prompt, it takes about 15 minutes to run, plus adjustments and checks, total around 40 minutes. After that, for books in the same