Knowledge management · Card-based approach ★ Coach Jiang's selection

Why Split Data into Cards? From PDF Readability to Card-Based AI Retrieval

A teacher gave me a package of materials, all PPTs and PDFs, with the same concept repeated in different files. To let AI take over this data, we first need data cleaning, then card-based approach. This page uses interactive animation to explain the entire logic clearly, and you can experience it page by page by scrolling down.

↓ SCROLL · Ten-page interactive · Approximately four minutes

What this article is about: PDF is a printed format, and AI reads it as coordinates and garbled text; the more data you give AI, the more the accuracy drops in a snowball fashion. The solution is data cleaning plus card-based approach: split knowledge into individual cards, use controlled tags to let AI retrieve the precise few cards, and recombine them like building blocks when needed. The following text is in the form of projected pages, each with an interactive or animation, and you can directly drag the slider or click the tags to try it out.

Who is this for
  • A pile of PPTs and PDFs in hand, wanting AI to help organize, but finding it can't understand the lecturer and knowledge workers
  • Have heard of the card box note-taking method, but don't know how it relates to AI
  • Those who want to turn years of accumulated old data into reusable material library
What you can take away
  • A directly copyable 'atomic decomposition' prompt (at the end)
  • Three specific reasons why PDFs are hard for AI to understand
  • Two accompanying items after card decomposition: controlled tag retrieval, 3X4 time-based storage
Act 01 · Current Situation

Your data is currently like this

PPT and PDF are created for 'teaching a class'. The same concept is taught in slide 12 of class A, slide 5 of class B, and again in the appendix of a PDF. Each time, the explanation is slightly different, but the core idea is the same.

Class A · p.12
Class B · p.5
Supplementary PDF · Appendix
↑ The three highlighted lines are actually the same concept

This is not wrong, teaching naturally requires repetition. The problem is that with more data, it's hard to determine how many times a concept has been taught, and which version is the best. Moreover, these files are very difficult for AI to read. The next page will explain why.

Act 02 · Why AI Struggles to Read PDF

PDF is a print format, designed for humans, not AI

The beautiful layout we see is actually code. The size, weight, font, and position of each character must all be written into the code. Even a single line separator must describe 'there is a line, with its origin at a certain point, drawn from one point to another'.

What you see

Four Principles of Teaching
PrincipleAsk the purpose first
TimingWhen students ask questions

What is actually stored in the file

BT /F1 14 Tf 72.0 704.2 Td (教) Tj 86.4 704.2 Td (學) Tj 100.8 704.1 Td (四) Tj 0.8 w 72 698 m 210 698 l S /F1 9 Tf 72.0 671.5 Td (原) Tj 81.0 671.4 Td (則) Tj 72 660 138 14 re S ...

A single cell in the table, which contains so many coordinates and stroke information. AI relies on 'coordinates being the same' to determine if it is the same row, so these things happen:

If the table does not match, it skips the row

If there is an accidental extra space or the coordinates differ slightly, AI will consider that data to belong to another row, even judging it as the next row, causing the data to become misaligned.

The text in the image

The text in the image cannot be read

The key points on the slide are often placed in the image, or the entire file is a scanned document. Those characters are pixels to AI, and there is a possibility of misreading them.

There is another point: The bullet points in a slide deck are prompts for the speaker. For example, you may speak for five minutes from a bullet that only says “1. Timing,” but those five minutes are not stored in the file. AI sees only the short label. Therefore, the first step is data cleaning: remove formatting instructions and garbled text, and retain the plain text. The simpler the container, the less interference there will be.

Act 03 · Why do we need to split

The more things you feed to AI, the more confused and scattered it becomes

AI's attention span is the same as humans. If you do not tell it the key points and feed it a lot of things, it will get confused and dizzy. The amount of data and focus are in a square ratio When the amount of data increases, the attention span and accuracy will decrease in a snowballing manner. Try moving the slider below.

Data fed to AI

Accuracy of AI

98%Precision
One cardNow feeding: 6 pieces of dataEntire data repository

Therefore, the data should be split into one card at a time. After splitting, the key is to let AI find the required few cards and process them How to find? Next page.

Act 04 · After splitting: Retrieval

More cards are not a problem, the problem is not being able to find them. Therefore, labels should also be made into cards

Splitting into cards is only the first half, the second half is letting AI find the precise few cards Method: Attach labels to each card, and the labels themselves are also cards, serving as the connection hub of the entire data repository. Try clicking on a label.

This is the completed card repository. When AI answers a question, it uses labels to narrow down the scope.

The context is like meridians densely interconnected, but you don't need to manage every single meridian. By grasping a few acupoints, you can control the entire meridian. Label cards are the acupoints.

Labels should be controlled and not invented on the fly

First create a controlled classification table, and select labels from the table. Assign coordinates to each card based on four dimensions: object, content, context, and state.

What if there are too many cards? Give them a time-based structure.

You can combine it with the 3X4 data organization method: content is divided into three types, and four time-based categories for storage. Frequently used items come to the top, and expired items sink to the bottom, preventing the card repository from becoming increasingly messy.

Act 05 · Cardification Four Principles

A qualified card looks like this.

This is the simplest application of the card box note-taking method: breaking down knowledge into individual cards, and recombining them as needed. The standard for breaking down has four criteria, and the complete prompt is attached at the end for direct copying.

1
Minimum independent unit

Each card contains a single complete concept. It is neither a chapter nor a class, but a complete idea that can be explained.

2
It can be understood independently, without relying on surrounding text.

When this card is taken out alone, it can be understood without flipping through previous or next pages. The bullet points should be written into complete sentences.

3
Mark the relationships between cards.

Mark which cards are related to this one. Later, when searching, you can find things through these relationships, as well as through the tags mentioned earlier.

4
Use Markdown to clearly define the boundaries.

Without a title, AI cannot distinguish where a paragraph begins or ends, and it will become a single block. With a title, the paragraph boundaries are clear, and both humans and AI can understand it.

Cardification is not for aesthetics; it is to enable knowledge to be truly reused.

Act 06 · Step 1: Break Down

Break a slide into several building blocks.

A single slide usually contains more than one item. Four bullet points may be four independent concepts that happened to be on the same slide. The act of breaking them down is to turn them into independent building blocks.

A Course · p.12 'Four Principles of the Teaching Context'
Card 01 Ask what he wants to solve first
Stuck 02 Do it once on the spot for the unclear parts
Stuck 03 Leave time for the feeling after demonstration
Stuck 04 Collect one sentence of feedback before class ends

One slide → four individual cards that stand on their own

Act 07 · Step 2: Merge

If the same concept is explained three times, keep only the best one

After splitting, you will find a lot of cards that look similar. Put them together, pick the best version, or merge them into a better one. Three cards become one, with the same content, but less repetition.

A class p.12
Ask the student what they want to solve before answering
Merged card
Common student questions often hide the real issue behind the problem. First ask, 'What do you want to use this for?' to uncover the real issue before answering.
Supplemental PDF appendix
Confirm his purpose before answering
Three cards → one card

Originally hundreds of pages of PPT, it may be condensed into dozens of cards. The number of pages decreases significantly, but the content remains the same.

Act 08 · Step 3: Reorganization

Blocks are usually scattered, and you need to assemble them into the required shape immediately when needed

This is the real value of cardization. AI can quickly call different cards to help you compose new content: writing a business plan, creating a new course, or replying to a question, all come from the same set of cards rearranged. Click the button below to see.

Six cards are here, and the sources were originally scattered across three different files.

Act 09 · Division of labor

The disassembly work is done by AI, and people only need to intervene in two places

The entire process does not require you to rewrite the data. Only two things truly need human involvement, because these two things only you know the answers to.

01

Repetition of the concept, you choose which explanation is best

The same concept has three versions. Which version has the best audience reaction, and which version you speak most smoothly, only you know. You choose, or you speak a better one, and AI writes it into a card.

02

Confirm whether the cards disassembled are correct

Some things in your mind are one thing, but AI may split them into two cards; some things you think are two things, AI may merge them into one. This requires you to explain, and AI will make the changes.

Act 10 · How to start

First, take the shortest one to try disassembling

STEP 01

Pick the shortest one

Pick the one you think is the simplest PPT or PDF, preferably within twenty pages. Do not pick the most important one; pick the one that is easiest to understand.

STEP 02

Let AI disassemble it for you to see

First, do data cleaning and convert it into plain text, then use the prompt at the end to split it into cards. After splitting, you will see a list: how many unique concepts are there in this file.

STEP 03

If you think it's correct, proceed with all of it.

Check whether the split result looks correct and whether the wording has deviated. If you think it's correct, use the same method to process the remaining data.

After organizing, creating new courses, writing articles, making online courses, and answering questions are all done by recombining the same set of cards. The material is processed once and then reused repeatedly.

Take away: atomic decomposition prompt

Directly input your current project, workflow, or article into AI, asking it to help you with 'atomic decomposition'. The prompt is ready, just copy and paste to use.

I currently have a [project/article/workflow] that needs to be organized. Please help me: 1. Decompose the content into 'minimum independent units', each unit containing one complete concept (That is, what is called cardization or atomicization) 2. Each unit should be able to be understood independently, without relying on the context 3. Mark the relationship between each unit 4. Convert it into Markdown syntax, with clear chapters and sections (Each chapter and section is a minimum independent unit) Content as follows: [Paste your content]

After decomposition, there are two complementary items, each with a depth article that can be read next: how to control tags and how to turn tags into cards, see <Turn labels into cards, and the knowledge base comes to life: Tag Wiki Linking Method〉.

★ Coach Jiang's selection Knowledge Management Card Box Note-taking Method Data Organization Teaching Article Zero-base entry