如何用AI学会所有东西:基于Obsidian+Claud (English)

Generated: 2026-06-23 13:43:25

---

You Should Never Manually Maintain a Knowledge Base! Does That Sound Like Heresy?

When I uncovered this truth last year, I was dumbfounded myself.

Since childhood, teachers have drilled into us: "A good memory is worse than a bad pen," right? You think you're learning, but what's actually happening? Most people won't even want to look at their own handwritten notes after three months. Don't believe me? Try digging up something you organized last year — I bet you won't even bother clicking the title.

Speaking of which, let me tell you about the pitfalls I've stumbled into.

A Perfectionist's Breakdown Diary

Last year, I followed a YouTuber's note-taking principles and, hyped up, jumped headfirst into the Obsidian rabbit hole.

Guess what I did? I photographed textbooks, used GPT to convert them into Markdown, and then — manually split everything into atomic knowledge points. I categorized, tagged, linked, never missed a step. I stared at the screen every day like a diligent little bee.

And the result?

Two weeks. Just two weeks. I crashed.

It was exhausting. Every time I added a new piece of knowledge, I'd agonize over where to put it, whether to update older pages, whether I'd missed any links. It was cumbersome, troublesome, and I felt like giving up at every turn.

Later, I wrote a script and used Claude Code to restructure the entire note vault. Felt amazing when it was done. But a few weeks later, it was useless again — the script only did a one-off job; new material wouldn't integrate automatically. I still had to maintain it manually.

See where the problem lies?

The very premise of "having a human maintain a knowledge base" is fundamentally flawed.

What should people be doing? Thinking, judging, asking questions — not acting as laborers. When you're constantly moving bricks, how can you spare the energy to envision what kind of house you're building?

Offload All the Grunt Work to AI — You Just Be the Boss

Once I realized that, I flipped my entire approach: AI continuously maintains the structure; humans just toss in raw materials and ask questions.

My current knowledge base, to put it simply, has just three layers. Don't be intimidated — let me explain in plain English:

Layer 1: raw (Raw Material Layer)

It's basically a trash bin. Web clippings? Toss them in. PDF papers? Toss them in. Podcast transcripts? Toss them in. I use Obsidian Web Clipper to snip content into a specific folder with one click. A keyboard shortcut, and external images are auto-downloaded locally — otherwise, URLs break within a couple of months, and you'll have nowhere to cry.

Layer 2: wiki (Wiki Layer)

This layer is the most important, but remember: humans only read, never write! Don't get itchy fingers. I couldn't resist at first either, always wanting to tweak things. Then I realized — you change something, AI doesn't know, and next time it compiles, it might overwrite your edits. They fight each other, and in the end, you're the one who breaks down.

So I set a rule: the wiki is AI's territory. Have an idea? Write it into the Inbox, or just give a direct command to the AI to make the change.

Layer 3: schema (Rule Layer)

It's just a single file that defines the rules of the knowledge base: structure, naming, processing workflows — everything clearly written down. This is the core asset.

And here's the thing: this file wasn't written all at once; it was gradually refined together with the AI. Discover the AI made a mistake? Go edit this file, and next time it'll get it right. A hundred times easier than manually fine-tuning a model.

You want to know how satisfying the actual workflow is? I integrated Claude Code via Obsidian's command plugin, so the entire vault becomes the AI's working directory. Toss new material into "raw," open it at night and say, "Compile the new stuff from Inbox into the wiki," and the AI runs on its own — reads the material, writes summary pages, updates all related pages, and appends a log entry.

You don't even need to watch the process. Wake up after a good night's sleep, and your knowledge base has already updated itself.

Think It's Expensive? You Might Be Surprised

Some people ask: Is the cloud API secure? Isn't it too expensive?

I use DeepSeek V3 together with Claude Code. The cost per million tokens is about one-tenth of Claude's, and the text comprehension is good enough. For a typical week's worth of clippings and compilation, it costs a few cents.

All files stay on your own computer. Obsidian is a purely local app. Of course, when the AI processes data, it goes up to the cloud API, but I only put non-sensitive materials into the Inbox. The boundary is clear: diary entries, project documents, etc., go into other folders and don't participate in the compilation.

More Interesting: The Knowledge Base Grows on Its Own

Did you fertilize it? Water it? No. But the garden gets denser every day.

Every time I add new material, the AI doesn't just build an index for you to search later. It actually reads it, extracts core information, and updates the existing wiki. If the new material conflicts with an old statement? It automatically marks it on the relevant page.

Was the previous conclusion correct? It reinforces it. Not correct? It overturns it.

With each added piece of material, the wiki becomes a little more complete. Cross-references are done automatically, contradictions are flagged. You don't need to re-derive everything from scratch every time.

When I was reviewing for finals, I tossed my midterm mistakes into the system. The AI automatically integrated them into the knowledge structure. I asked questions using the Socratic method, and the AI answered precisely based on the wiki, even pointing out where my understanding was shallow. Way faster than flipping through my notes.

But There's a Big Trap — You Must Avoid It

The tool world churns out new things every day: OpenClaw, Hermes Agent, all sorts of "skil" — everything looks useful. It's easy to invent needs to use them, or to fear that others will use them and leave you behind.

The result? You haven't even mastered the tools, and your time is already wasted.

My experience is simple: Start from your current pain point, not from the tool.

I didn't build this system because Claude Code became popular. I did it because my paper notes were a mess I couldn't find anything in, and my photo archives were useless too. I was forced to find a solution. Tools are the solution to a need, not the other way around.

If your note volume is small and you can find things by searching, don't bother with this setup. Wait until your knowledge base is so large that "relying on memory can't find things," then build it. And the good news is: setting it up from scratch takes about two hours, and the first article can be in the wiki within 15 minutes.

Two Steps You Can Start Right Now

First, throw in your first article. Use Obsidian Web Clipper to clip an article into the designated folder. Then open your terminal and run the command: "Process this article according to the rules." The AI will read the article, build a summary page, and update the index.

Look at the result. If you're not satisfied, have it revised, and write the feedback into the rule file.

Don't forget to note the current version number. Three months later, look back — you'll be amazed at what it's become.

The essence of knowledge management has never been storage; it's connection. In the past, we relied on manual linking. Now, AI does it. Your job isn't to be a switchboard operator; it's to decide which wires should connect and where —

Then ask the AI: Did this path work?