Solving Problems in the AI Era

It has been a while since my last post. I have called this article “Solving Problems in the AI Era.” That is where I will begin, but I will not limit myself to it. This is also a look back at a winter break spent experimenting, figuring things out, and thinking along the way. I hope some of it proves useful or sparks an idea. I would love to discuss it with you.

Overview

Approaching unfamiliar problems: How should we deal with new problems in the AI era?
Misunderstanding AI’s capabilities: When working with text and images, if AI seems incapable, chances are you have not learned how to use it effectively.
Building a personal knowledge base: Using Claude Code and Obsidian to create an intelligent knowledge base.
Thoughts on OpenClaw and the future of products.

1. How Should We Approach New Problems?

Let me start with a concrete example.

Take document formatting. The traditional approach is to format a document, such as a paper, manually. That might take half an hour or as long as two hours. A more advanced approach is to use existing tools such as Microsoft Office or WPS. But in my own tests, the results were disappointing. WPS offered only a few preset formatting options, which could not meet more specific requirements.

Even this more advanced approach fell well short of what I wanted. You might ask, though: Doubao and DeepSeek cannot handle that formatting task either, can they?

Here is the point I want to put forward:

When working with text and images, if you think AI is not capable enough, chances are you have not learned how to use it effectively.

I will explain that claim in more detail below. In this particular case, I eventually found that I could equip Claude Code with a docx skill, giving it the ability to inspect and manipulate docx files. I then described what I needed in the terminal, and it handled the formatting itself. In practice, the results were very good.

When we encounter a new problem today, we can choose to extend the capabilities of an AI tool we already use, such as Claude Code, so it can handle the task. Better still, the benefits compound: as you keep using it, you can make it increasingly capable.

Why I Make That Claim

I specify text and images because current AI capabilities largely derive from foundation LLMs, whose core strength is language. Taking some models’ multimodal capabilities into account, the areas they excel in are, for now, primarily text and images. That still covers the vast majority of what people want to do with AI.

My view is that when the underlying model is the same, differences in results mainly come down to two factors: context and skills.

Context: Prompts and the Working Environment

Context includes your prompts and the working environment. Prompts need little explanation: the more specific your requirements, the better the results.

If you have mostly used chatbots such as Doubao or DeepSeek, the idea of a “working environment” may be less familiar. It is also a key part of the move from chatbots to agents.

Consider Claude Code, a familiar example of an agent:

It operates in a folder on my computer, while the model computation still happens in the cloud.
It can access my local files.
This is like a more advanced version of uploading files to Doubao: you can provide many files and expose the folder structure so the AI can find and read what it needs.
It can do more than read. By running command-line operations on your computer, it can modify files for you.

The widely discussed OpenClaw is essentially this kind of agent too, with further extensions to its capabilities. I will share some thoughts on it later.

Skills: Extending What AI Can Do

Beyond context, skills are another key factor in the quality of AI’s work.

Although an AI agent can run command-line operations on your computer, that alone has limits. If you want it to do a particular kind of work, such as manipulating docx files, you need to install an appropriate set of instructions and tools, commonly called a skill.

As you use the agent, you can keep adding skills according to your needs, making it more capable and better suited to your own workflow.

A Practical Example

I followed these principles to write a summary of our fieldwork team’s activities:

I supplied all the interview transcripts, visit plans, and other documents from our winter-break fieldwork as context.
Over several rounds of conversation, I gave the AI precise and comprehensive instructions.
I equipped it with a skill for working with docx files.

In under 10 minutes, I had a summary of more than 3,000 Chinese characters. It was grounded in what we had actually done, rather than sounding like an empty piece invented by AI. I think the high-quality context I provided made a large difference.

On Writing That “Sounds Like AI”

As for the familiar “AI flavor,” I think it is simply what happens when AI lacks context.

Imagine being asked to write a fieldwork summary without being told anything about the fieldwork. Could you write a good one? Under the same conditions, AI produces the writing people describe as sounding like AI.

But if you provide rich context, it is almost as if the AI took part alongside you. It can then produce an excellent summary without that generic AI flavor.

2. Building a Personal Knowledge Base with Claude Code and Obsidian

The discussion above leads naturally to the next question: How do we give AI enough comprehensive context?

My answer is to build a personal knowledge base.

Why a Knowledge Base Matters in the AI Era

Such a knowledge base has a special role today: it is not just for you to read. It is also for AI.

It can help AI understand:

What you have been doing recently.
Good articles you have read.
Your own thoughts and reflections.

When AI responds with that background in mind, its answers can become much better and resonate more deeply with you.

Some Everyday Uses

For example:

I can organize my class notes there and ask AI to answer questions based on what was actually taught.
If I forget a course requirement, such as a homework deadline or how grades are calculated, I can simply ask, provided I have recorded the relevant information.

How I Organized My Knowledge Base

After several days of experimenting and tinkering, I finally finished the initial setup. I will continue improving it.

The core files are:

AGENTS: My instructions for AI.
INDEX: A guide that helps AI locate files quickly.
MEMORY: AI’s long-term memory of me, maintained by the AI itself.
USER: A profile of me as the AI sees me, maintained by me.

Returning to a Daily Journal

I have also started keeping a daily journal again.

One inspiration was an article by Kashen titled, in translation, “Why I Strongly Recommend Keeping a Journal in the AI Era.” I had been thinking about this: once AI became a capable assistant, it could complete much of my work without me doing it by hand. Yet even when I supplied the instructions, how much of me remained in that code or that article? What made it mine at all?

The answer is simple: I remember. My experiences, preferences, and sleepless nights form a continuous story. That story is me. If it were removed and replaced with someone else’s story, I would no longer be myself.

Liu Heng once wrote, in translation:

“After artificial intelligence can replace almost every form of human expression, what alone survives is each individual’s unique expression of their own life.”

I want to preserve that story and that unique expression, starting by recording each day’s progress and discoveries.

The other reason connects to the original purpose of the knowledge base: I want AI to understand what I have been thinking about and doing lately. Journaling serves both purposes.

3. Thoughts on OpenClaw and Future Products

I began following OpenClaw when it first appeared, back in its Clawdbot days. To be honest, because I was already using Claude Code, it did not strike me as revolutionary.

It did open up ways for AI to interact with people through instant messaging, connecting to services from Telegram to Feishu. AI could finally interact with people through their messaging platforms. There were also ideas such as heartbeats and SOUL.md. But at this stage, I do not see enough added value to feel anxious about keeping up.

Its founder has also openly acknowledged existing security issues and difficulties with setup. Even as major companies announced their own initiatives, I did not invest much effort in it. I already had Claude Code, and for now that was enough.

Much of the media hype sells anxiety. After spending so much time around AI, I have become fairly immune to it.

Despite the hype and restlessness, OpenClaw’s breakout popularity has had real effects:

It has expanded the public’s sense of what AI can do beyond chatbots.
It has generated substantial revenue for model companies and cloud providers, convincing investors that AI can make real money now rather than being only a bubble.

Let the bubble keep growing!

Where Products Might Go Next

To finish, a few thoughts on the future direction of products.

1. A Tool AI Cannot Use Is Not a Good Tool!

Obsidian’s introduction of a CLI has only strengthened this belief for me.

I think all products will need to move in this direction, unless they deliberately keep AI out—as I suspect WeChat will. AI will become the power user of utility software. If AI cannot use your product, you lose your largest potential user, and decline becomes likely.

This is also a major opportunity. I expect an instant messaging product capable of rivaling WeChat to emerge soon, with one defining feature: AI can use it. Feishu and DingTalk are among the candidates.

Imagine AI helping you handle the messages and notifications across your social platforms, identifying what needs your attention and what you can ignore. That would be a radically different experience.

2. If Your Product Is for People, Try Adding a Chat Interface!

People accustomed to chatbots will be less willing to hunt for information or perform operations themselves. The new pattern is to make a request in a conversation and let AI act or search on their behalf.

Forums, for example, should experiment with this approach. The task would shift from “find a thread” to “have AI find relevant threads and summarize them.”

It has been a long time since I wrote an article this long myself. When I keep AI out of certain tasks, such as writing, it feels primitive—but also pure, grounded, and satisfying.

Thanks for reading!