# What Helping My Son Win an AI Agent Award Taught Me About AI Education

Willdom Lee · 2026-04-18

[HTML](https://willdomlee.com/en/articles/ai-agent-contest) · en

Over the Spring Festival holiday, I helped my son prepare for the AI Agent Design and Application category of an information technology innovation competition organized by the education bureau. The results recently came out, and he won a small award.

I heard that the judges thought the project was good and that it could have done better had we understood the brief correctly. That was a little disappointing, but my son and I both saw it as a lesson for future competitions. He was happy to win, and I wanted to record a little of the process, along with a few thoughts on school education. First, here were the competition requirements:

![](https://willdomlee.com/assets/a/ai-agent-contest/02.webp)

This was the first year the competition included agent design, a sign that Shenzhen's education system is beginning to take action on AI awareness and practice. The requirements were fairly standard, covering students from primary school through university, with different expectations at each level. Here were the primary school requirements. Pretty simple, right?

> 1. Abstract and keywords
>
> Abstract: Explain in one paragraph what you made, how you made it, and what it achieved.
>
> Keywords: Three to five terms, such as “agent,” “large language model,” or “community service,” chosen to suit your project.
>
> 2. Research question
>
> State the question you are investigating in one or two sentences.
>
> 3. Background and significance
>
> Explain in everyday language why the question is worth exploring.
>
> 4. Observations and data
>
> Use tables and charts to present your observations and analysis. Include units and the dates when the data was collected.
>
> 5. Conclusions and improvements
>
> Summarize what you found, what it means, and one or two things you would improve next time.
>
> 6. References (optional)
>
> List relevant references with their titles, sources, and dates.

We had finished building an agent during the Spring Festival break, before the detailed brief was released. When it came out, we realized our project had missed the point. We hesitated: should we start over? It was my first time helping him build an agent, and getting started had taken quite a while. In the end, we agreed to rebuild it around the new rules. The good news was that he had already learned the basics and could work more independently the second time. Unfortunately, we still overlooked one requirement, which led to the disappointment I mentioned earlier.

The first thing I want to say is that a child's interest matters enormously. If he had not been interested, and still did not want to take part after some encouragement, we would simply have stopped.

As parents, we need to draw clear boundaries when helping our children. We can guide them, provide a better environment for creating things, and answer questions along the way, but we must not do the work for them. We have always treated winning as the second priority; practice comes first. Building an agent lets children explore the strengths and limitations of different models: what model size means, what a chain of thought is, and what multimodal models actually do. They can also see that a model acting as a brain needs hands, feet, ears, and eyes to solve problems in specific situations.

Take our project as an example. My son made an “AI Grandson” for his grandmother, with four main functions: weather reports, outfit recommendations, health management, and recipe suggestions. Personalization was the point. Weather reports used her location; outfit suggestions combined the weather with clothes she actually owned; unusual health readings could trigger a background notification to Dad's phone; and recipe suggestions could use a photo of the fridge as input. Here is the mind map he made:

![](https://willdomlee.com/assets/a/ai-agent-contest/03.webp)

We used ByteDance's Coze programming platform to build the agent, and found it very easy to use. The overall setup was:

- The brain: the Doubao 1.8 model
- The personality: a prompt
- The skills: Moji Weather, time and location lookup, text-to-speech announcements, and messages to Dad. These gave the agent its hands and feet.
- The memory: a memory store for Grandma's clothes, health data, and other information

![](https://willdomlee.com/assets/a/ai-agent-contest/04.webp)

While building it, my son tried several models, comparing both answer quality and response speed—speed matters to older users. I took the opportunity to explain their different strengths and why those differences existed. Hallucinations became a particular focus during testing, and we worked together on ways to eliminate them. These experiences help a child see that models and AI are not all-powerful. That matters.

I found that even primary school children take naturally to prompt engineering, because the material is something they can read and understand. After changing a prompt, they can immediately see the result in the playground on the right side of Coze. It does not get much more direct than that. I also found that a child can move seamlessly from prompt engineering to vibe coding. Lately, my son has been using Coze to make a racing game he enjoys, recreating the North–South course from Pegasus 3. We can even race each other when we have a spare moment. He has never learned to code, but that does not stop his imagination from running free. That is the appeal and the point of AI.

When testing speech synthesis, he adjusted parameters to try different voices. The first time he heard lifelike speech generated from text, he was amazed. Those little moments of amazement, brought about by a child's own hands, give their imagination more room to grow.

![](https://willdomlee.com/assets/a/ai-agent-contest/05.webp)

Hands-on work matters. To make the AI Grandson better, I suggested he do two things:

1. Interview Grandma carefully and take notes in an exercise book. Learning more about her would give him a solid foundation for a thoughtful, personalized agent.

2. Use his phone watch to photograph the clothes she wears regularly, so the daily outfit recommendations would be grounded in reality.

![](https://willdomlee.com/assets/a/ai-agent-contest/06.webp)
![](https://willdomlee.com/assets/a/ai-agent-contest/07.webp)

When the agent combined that day's weather with the collection of clothing photos he had taken on his watch and suggested a suitable outfit, we had another moment of amazement. I explained roughly how it worked and gave him a simple introduction to multimodal models.

![](https://willdomlee.com/assets/a/ai-agent-contest/08.webp)

The free tokens from Coze were not enough, so we bought a 9.9-yuan monthly plan. I also explained why Coze charges money and what resources model inference consumes behind the scenes.

Finally, he drew a mind map on a whiteboard from IKEA and recorded an explanation on his own; I just opened Tencent Meeting for him. I edited the video and added subtitles in Jianying, then exported the agent and sent the package to his teacher. Because he had built it himself, he could explain it without prompts and finished the recording in two takes. That exceeded my expectations.

![](https://willdomlee.com/assets/a/ai-agent-contest/10.webp)

Two things we had done previously also seemed to help him understand AI and how agents are put together:

1. Reading books about AI, including Introduction to Artificial Intelligence: Primary School Edition, edited by Ma Shaoping and published by Tsinghua University Press.

2. Learning basic programming with a RoboMaster S1 programmable robot. See [Exploring Robot Programming with AI and My 10-Year-Old Son: A Father's Experiment in Learning](https://willdomlee.com/en/articles/learning-with-my-son.html).

![](https://willdomlee.com/assets/a/ai-agent-contest/11.webp)

I have always believed learning should not be too narrowly goal-driven. When we spend time studying something seriously, that knowledge will find a way to resonate with our lives eventually.

I wrote this casually, without AI assistance, polishing, or proofreading. It may not be entirely rigorous, and it is rather chatty. Take it as it is—and if it sparks a thought, all the better.

---

Originally published in Chinese on WeChat, 南山独唱团. [Read the original](https://mp.weixin.qq.com/s/R58_jD7d7mm-ewMCUsaSxw).