# Bringing AI into the Classroom: Building an OpenClaw Assistant for a Homeroom Teacher

Willdom Lee · 2026-02-24

[HTML](https://willdomlee.com/en/articles/ai-in-the-classroom) · en

## 1. A busy homeroom teacher at home, and a reason to try OpenClaw

There is a lovely teacher in our household who currently serves as a middle school homeroom teacher. Anyone who has been to school knows how demanding that role can be. Ms. Zhang takes her work seriously. Alongside looking after several dozen lively children, she moves between group chats and spreadsheets, analyzes exam results, tallies classroom points, and carefully chooses her words when communicating with parents. Each week brings a few unexpected situations too. My description hardly does it justice; fellow teachers will understand the effort involved.

After trying an OpenClaw deployment, I had a reasonable sense of its capabilities and uses. I realized that a homeroom teacher could really use an AI assistant—a deputy always ready to help. I decided to build one on OpenClaw that could connect isolated data, understand education, and work autonomously. I called it Ban Xiaozhi. I began with a few things I considered essential:

- An in-depth interview with Ms. Zhang about her classroom management and educational philosophy. This would shape the assistant's values, approach to education, experience, and skills.
- Data cleaning, followed by building a database and knowledge base. High-quality data would help the assistant support classroom management more efficiently and personally.
- Hardening the server's security configuration. Data security matters.

Here is how it unfolded.

## 2. The architectural principle: make technology fit the work

Having already experimented with OpenClaw, the technical side was familiar this time. Here is a brief overview of the architecture.

There was no need to buy a Mac mini; I rented a cloud server instead. I chose Tencent Cloud Lighthouse, which offers one-click OpenClaw deployment and quick integration with four major Chinese messaging platforms: WeCom, QQ, DingTalk, and Feishu. This is not an ad. A four-core server with 4 GB of RAM and 3 Mbps bandwidth cost 79 yuan a year—excellent value, though model API usage is extra. After comparing several models, I chose Kimi for its stability and quality.

![](https://willdomlee.com/assets/a/ai-in-the-classroom/01.webp)
![](https://willdomlee.com/assets/a/ai-in-the-classroom/02.webp)

For data, I separated frequently updated records from relatively stable information.

Hot data lives in Feishu Base: frequent, small updates such as attendance and discipline records. This preserves Ms. Zhang's habit of quickly recording things on her phone. Connecting OpenClaw to Feishu Base lets AI read the data directly. Its access is read-only; it cannot change those records, protecting consistency and security.

Cold data lives locally in SQLite. Basic class information and results from major exams are stored in a local database for data security.

The original data also had familiar problems: fragmentation, weak relationships between records, inconsistencies, and duplicates. Using it as-is would have caused trouble, so I spent time designing the structure and cleaning it up. Preparation pays off, especially when AI can help with the cleaning. Here is the resulting structure:

![](https://willdomlee.com/assets/a/ai-in-the-classroom/03.webp)

OpenClaw's first successful connection to Feishu Base was a striking moment.

![](https://willdomlee.com/assets/a/ai-in-the-classroom/04.webp)

## 3. Giving the assistant a character: turning educational principles into skills

With the architecture in place, I reached the most important part: defining Ban Xiaozhi. There is nothing mystical about it. It means writing a few text files clearly and thoroughly. But it does take serious thought, which is why I interviewed Ms. Zhang first. If you are experienced in a field, you may already have much of this knowledge. The best prompt engineering is really the careful distillation of domain expertise.

These OpenClaw Markdown files form the core of its AI skills.

![](https://willdomlee.com/assets/a/ai-in-the-classroom/05.webp)

|  |  |
| --- | --- |
| **File** | **Main purpose in one sentence** |
| USER.md | Defines Ms. Zhang's working habits, communication preferences, and expectation that AI should serve as both an abacus and a radar. |
| IDENTITY.md | Sets the AI's basic persona, its role as an all-round homeroom assistant, and its overall communication style. |
| SOUL.md | Establishes the AI's core values and firm rules for emotionally intelligent communication. |
| TOOLS.md | Maps the underlying hot and cold data sources, tool-use conventions, and strictly read-only access boundaries. |
| AGENTS.md | Defines the AI's daily startup procedure and proactive duties, such as finding insights across data sources and prioritizing action requests. |
| HEARTBEAT.md | Configures scheduled background tasks, giving the AI a regular pulse for reviewing conversations, consolidating memory, and preparing daily briefings. |
| MEMORY.md | Provides a long-term memory archive for important student behaviors and events automatically distilled from everyday conversations. |

One example skill file: IDENTITY.md

![](https://willdomlee.com/assets/a/ai-in-the-classroom/06.webp)

Once these definitions were clear and combined with the connected data, Ban Xiaozhi's answers improved dramatically. For a systematic guide to writing skills, see my earlier article [How to Build Portable AI Skills Across Platforms](https://mp.weixin.qq.com/s?__biz=MzA4MDc1MjUxOA==&mid=2247484055&idx=1&sn=49143d7381ec9c9d4ea4b1bb09804624&scene=21#wechat_redirect).

## 4. What the assistant can do: the screenshots speak for themselves

The data below was created for demonstration and is not real student data.

### Combining multiple exams to analyze results and recommend priorities

![](https://willdomlee.com/assets/a/ai-in-the-classroom/07.webp)

### Attendance analysis connected to Feishu Base

![](https://willdomlee.com/assets/a/ai-in-the-classroom/08.webp)

### Creating slides and invitations for a parent meeting, with the teacher refining the first draft

![](https://willdomlee.com/assets/a/ai-in-the-classroom/09.webp)

### Making birthday cards for students born in February and sending them to me as a package

![](https://willdomlee.com/assets/a/ai-in-the-classroom/10.webp)

### A back-to-school reminder to get everyone focused again—something I used to think was not particularly useful 😄

![](https://willdomlee.com/assets/a/ai-in-the-classroom/11.webp)

There are many more examples, but I will stop here. If it develops in a particularly impressive way, I will share more when time allows.

## 5. Reflections and realizations

After putting the system together, I found myself thinking about a few possible futures:

- A personal team of experts for AI-empowered individuals: knowledge workers may no longer have to struggle to adapt to bloated, generic software. With agent frameworks, any person or organization can build a team of specialists that fits their own workflow.
- The rise of language user interfaces, or LUI. AI also makes command-line tools smarter, opening new possibilities for CLI interaction. A single instruction can have AI query and join tables, create slides, or even send email. OpenClaw connected to Feishu Base remarkably quickly, thanks in large part to Feishu's openness and AI-native design. Product creators should take note: products must be easy for AI to use as well as people, through APIs and interfaces that do not need a display. That trend is inevitable.
- A return to education's purpose: the stronger AI becomes, the more the human role in education matters. When AI takes over low-value tasks such as adding scores and formatting documents, teachers can spend more time helping children grow. After all, the jobs those children will do may not have been invented yet.
- A golden age for AI architects: people who understand a business deeply, uncover its needs, know AI, and can solve engineering problems will be exceptionally scarce.
- Rethinking management: as AI joins teams as a kind of silicon-based life, traditional disciplines such as leadership and business management will have to be rewritten. I took another look at the leadership book on my shelf. Its author clearly had not considered how to lead tireless AI toward a goal. Interesting times.
- Greater security challenges: the growth of agents will bring a vast number of security issues. I once asked OpenClaw to control our home printer directly. The moment it started printing was impressive, but it also sent a slight chill down my spine. From physical safety to data security, the tests are only beginning.

What we can do now is embrace AI and work hard to make ourselves and our organizations AI-ready. That takes more than a change in mindset. It requires sustained work on three things: **mature, continually evolving core business processes; high-quality private knowledge bases; and continuous data governance.**

Wishing everyone a good start back at work!

Security notes

API key security: keep model API keys private. Never expose them in public code or third-party images, where a leak could lead to unauthorized charges. Set daily spending limits and billing alerts to control the risk at its source.

Access control: follow the principle of least privilege when granting OpenClaw permissions. For local execution, allow access only to necessary folders, not unrestricted read/write access to the whole drive. For messaging integrations, enable only the required message and event permissions. Do not blindly grant everything, which risks data leaks or malicious actions.

Containing execution risk: beginners should test on a cloud computer or virtual machine, rather than granting broad access on their primary computer. This helps avoid irreversible damage from mistakes such as deleting files or changing system settings. Start with low-risk tasks, such as managing a schedule or drafting emails, and expand permissions gradually after checking reliability.

---

Originally published in Chinese on WeChat, 南山独唱团. [Read the original](https://mp.weixin.qq.com/s/RB0XesQzqWx_9bePQOuo-A).