2025.08.03 Lesson learnt from Vibe Coding to the Kids

In Korean

금일은 Vibe Coding으로 11~15세 아이들에게 웹게임을 만드는것을 재능기부로 강의했다. 뭐 시중에서는 인당 30만원씩 받는다지만, 회사에서 버는 돈으로 충분하다. 안그래도 학원을 다니며 살아가는 한국아이들의 삶에 부정적인 시각을 가지고 있는 내 신조가있어서, 나는 이런것으로는 돈을 안받는다. 받더라도 캄보디아 희망학교에 기부할예정.

아이들에게 전달하고싶은 가장큰 메세지는 “Questions are all you need”. AI를 통해 조금 더 비판적이고, 비평적이고, 의심하는 메카니즘을 머리속에 심어줘서 끊임없이 질문하여 방향을 잡아나가는것을 알려주고싶었다.

LLM을 쓰는 인간의 흔한모습은 “OO해줘, XX찾아줘, ??은 왜그런거야” 등이다. 이것은 단순히 LLM을 보조도구로만 쓰는 형태에서 더이상 발전이 없는 모습에 전형적 모습이지만, 개발자를 포함한 대부분의 사람들이 이렇게 쓴다.

최근 B-Factory 노정석대표님의 팟캐스트에서 보고 나도 꽤나 열심히 떠들고다니는 개념이있다. 많은 Big tech 혹은 Frontier Model을 만드는 곳의 리더들의 팟캐스트들을 보다보면 꽤나 많이나오는 “Capability Overhang” 혹은 “능력과잉” 이라는 개념인데, 이미 Frontier Model들은 이미 인간의 능력으로는 알지못하는 지평선 너머에 그 능력이 있는데 이를쓰는 인간의 능력 특히 “질문”하는 능력이 떨어지기 때문이라는것이 주된 내용이다.

LLM은 태생적으로 인간의 질문을 유추하여 인간을 만족시킬만한 답변을 내도록 되어있기때문에, 지금의 Frontier Model들은 기본적으로 인간이 질문한 정도의 데이터를 기반으로 그럴듯한 답변을 줄수밖에없다.

멀티모달기법들이 엄청나게 발전하고있어서 Input의 throughput가 늘면 조금 더 질문자의 의미를 전달할수 있겠지만, 언어로만으로 전달할수있는 지금으로서는 질문수준은 채팅으로 전달할수있는 아주 간략한 “Vibe” 즉 느낌정도가 다 인것이다.

강의를 목적을 달성하고자 질문자도 잘모르는 질문자의 “Vibe” 즉 느낌적인 느낌을 글로 표현한다는 것은 성인들도 어렵다. 실제로 이를 생각해내는것 조차 강의를 듣는사람들은 에너지 소모로 느끼는것이 원인인지 대부분은 “OO해줘” , “그거말고” 이런식으로 AI 와 대화하는걸 대부분본다. 사실 나도 가끔은 그렇고..

이 방법에서 벗어나기위해서는 뭐 어려가지 기법이있지만, 처음엔 간단한 컨텍스트를 줘서 질문하고, 그 답변을 가지고 다른 AI LLM에게 가져다줘서 비판적으로 생각하게하여 나도 같이 생각하고, 또 다른 AI에게 이 피드백을 주면서 또 같이 공부하는, Multi AI Persona기법이 내겐 가장 좋고 맞더라. 특히나 아이들에게도 더 좋을것같았다.

이번 강의의 샌드박스를 구상할때, 일단 게임이라는 흥미를 유발하는 아이템으로 시작하니, 집중할수있는 시간이 30~40분이 한계인 나이의 아이들이 2시간을 넘어서 밥먹는 시간도 아까워할정도로 열심히할수있도록 만들었다.

물론 준비는 Claude Code 와 Gemini Research를 통해 단일 웹페이지에서, 강의에 필요한 모든 Sandbox가 Win/Mac가리지않고 Oneline installation되도록 해놨다.. God CC…

Persona기법을위해 Cursor의 창을 열고 우측의 Agent창에게 인격을 부여하도록했다. 아이들이 각자 자기소개를 Agent에게 하고 인격을 부여하며 미리 만들어둔 Cursorrules와 여러 Markdown 문서들을 읽어들인Longterm memory들로 부터 이제 게임을 만들어주는 마법사로 변한다.

이제 부터는 아이들의 상상력을 발휘하는 시간. 처음에는 아이들이 “선생님 이건어떻게해요? 저건어떻게해요?” 하면서 내게 직접물었지만, 내답변은 확고하다 “그 질문을 마법사에게 물어봐”. 그리고 AI가 자세히 답변해주면서 Cursor에 내장된 AI Agent가 그거에대한 질문자의 의도를 파악하여 일을 진행하고, 그 코드가 Live server extension으로 연결된 브라우저로 바로바로 반영되니 아이들이 너무너무 집중하고 흥미로워하더라.

기억에 남는 재미있는 일화가 있다. 의도적으로 Short term memory를 리셋하고 새로운 에이전트에게 여태한일을 정리하여 넘기는 것을 알려주려고, “자 이제 너와 헤어지고 다른 AI 도우미에게 일을 맡겨야해. 우리가 했던일을 정리하고, 그 친구에게 전달할 말들을 만들어줘” 라고 시키라고했다. 그러자 아이들은 첫번째 에이전트 컨텍스트 윈도우에서 일했던 마법사와 헤어지는 것을 힘들어하더라. “왜요 ? 그럼 여태까지 한일은 어떻해요? 안헤어지면 안돼요 ?” 뭐 이런 질문들을 하는것이 얼마나 귀엽던지 ㅋㅋ AI 페르소나를 쓰게되면 자꾸 의존하게 된다는 연구결과를 봤었는데 이걸 아이들에서 바로 느끼게 된 사례다.

그리고 두번째 AI Agent에게 내가 만든 게임을 평가하라고 맡기는것을 시키라고했다. 아이들은 “얘가 뭘 안다고 제 게임을 알아요?”, “전 첫번째 애랑 할래요!!” 뭐 이런 재밌는 답변 ㅋㅋ 아이들이 자연스럽게 두번째 Agent로부터 비판적인 시각을 프로세스하도록 유도한것이니 처음엔 불편할 수밖에없다. 하지만 이런 과정을 통해 결국 질문을 잘하는 수준에 도달할 것이니.. 열심히 가이드하고 자리에서 알려주니 금방 두번째 에이전트에도 적응했다.

이정도 프레임워크를 만들어놓으니, 그 다음부터는 사실 내가할일은 아이들이 만든 게임을 Github page로 퍼블리시하는 파이프라인을 만들어주도록해서 선생님으로서 아이들의 게임을 플레이하고 즐거워하고 피드백주는 일이 다더라. 얼마나 즐거워하고 몰입하던지..

현업으로 바쁜와중에 준비하는 과정이 녹록치 않았지만, 일단은 너무 보람찼다. 이러한 교육의 프레임워크는 앞으로 우리가 AI와 함께 살아가는 현시대에 필요한 과정중에 하나가 아닐까 싶기도하고…

In English

Today I donated my time with Vibe Coding, showing 11- to 15-year-olds how to build a web game. The going rate is about ₩300,000 per kid—money plenty of tutors gladly scoop up—but I have no desire to cash in on parents’ ambitions. My salary covers my needs, and if I ever do accept a fee it’ll go straight to the Cambodia Hope School.

The core message I wanted to plant was simple: “Questions are all you need.” I tried to wire a critical, skeptical mindset into their heads—nudging them to keep probing, steering, and re-shaping ideas with AI as their sparring partner.

Most people, even developers, use LLMs in a shallow way: “Do X,” “Find Y,” “Why is Z like that?” It’s the default “assistant-only” posture that never evolves. Lately I’ve been riffing on a concept I heard on B-Factory CEO Roh Jung-seok’s podcast in the youtube: capability overhang. Frontier models already hold skills beyond our grasp, but we can’t unlock them because we humans are lousy at crafting questions.

LLMs are wired to infer intent and give answers that fit the prompt, so they can only rise to the level of the question they’re fed. Even with multimodal inputs widening the channel, a text prompt still boils down to a faint vibe—a sliver of intent squeezed into a chat bubble.

Expressing that vibe is hard even for adults. Learners quickly feel the mental strain, so they slide back to “Do this” or “No, not that.” (I fall into that rut, too.)

My escape hatch is a Multi-AI Persona workflow: give one agent a brief context, pass its answer to a second agent for critique, chew on the feedback together, then hand it to yet another agent. Rinse, repeat. It works beautifully for kids.

For today’s sandbox I used a game as bait. Kids who usually tap out after 30 minutes stayed locked in for two hours, even resenting the lunch break. Claude Code and Gemini Research let me bundle everything into a one-line install on any machine—chef’s kiss.

Inside Cursor, each child spun up an agent, introduced themselves, and loaded my pre-written cursor rules and Markdown guides. Their AI companions instantly morphed into personal game-building wizards.

Hands shot up: “Teacher, how do I do this?” My answer never changed: “Ask the wizard.” The agent parsed their intent, wrote code, and Live Server pushed updates straight to the browser. The room went silent—total absorption.

One highlight: I wiped an agent’s short-term memory and told the class, “We have to hand the project to a new helper. Summarize what you’ve done and craft a message for the next AI.” Panic! “But what about our work? Can’t we keep our wizard?” Adorable evidence that humans bond with AI personas.

Next, I made them ask a second agent to critique their game. Grumbling erupted: “What does this new guy know?” Exactly the discomfort I wanted; they were learning to welcome critical feedback.

With the framework humming, my role shrank to wiring a pipeline that publishes their games to GitHub Pages, play-testing, and cheering them on. They were so absorbed I had to pry them away from the keyboards.

Preparing all this while juggling my full-time job wasn’t easy, yet the payoff was huge. Frameworks like this—teaching the art of questioning and collaborating with multiple AI personas—feel like essential life skills for the age we’re entering.

And that’s a wrap.

나의 말:

학원시장에서 30만원쯤 한다는것으로 부드럽게.. 지금은 너무 공격적이야 표현이

ChatGPT의 말:

6초 동안 생각함

Today I volunteered with Vibe Coding, guiding 11- to 15-year-olds through building a web game. In the private-academy scene a class like this typically runs about ₩300,000 per student, but my day job already covers my needs, so I’d rather share the knowledge freely. If I ever do accept a fee, it’ll head straight to the Cambodia Hope School.

The core message I wanted to plant was simple: “Questions are all you need.” I aimed to wire a critical, skeptical habit of mind—constantly probing, steering, and refining ideas—with AI as their sparring partner.

Most people, even developers, interact with LLMs in a shallow way: “Do X,” “Find Y,” “Why is Z like that?” It’s the default “assistant-only” pattern that doesn’t evolve. I keep referring to a concept I heard on B-Factory CEO Roh Jung-seok’s podcast: capability overhang. Frontier models already possess abilities beyond our horizon, yet we can’t unlock them because we’re still clumsy at asking questions.

Since LLMs can only rise to the level of the prompt, our modest queries limit their answers. Even with multimodal inputs widening the channel, a text prompt still boils down to a faint vibe—a sliver of intent squeezed into a chat bubble.

Expressing that vibe is tough even for adults, so learners quickly slide back to “Do this” or “No, not that.” (I catch myself doing it, too.)

My workaround is a Multi-AI Persona workflow: give one agent a brief context, pass its answer to another for critique, think together, then feed the feedback to a third. Rinse and repeat. It works beautifully for kids.

For today’s sandbox I used a game as the hook. Kids who usually tune out after 30 minutes stayed locked in for two hours, barely wanting to break for lunch. Thanks to Claude Code and Gemini Research, everything installed in one line on any machine—chef’s kiss.

Inside Cursor each child spun up an agent, introduced themselves, and loaded my prewritten rules and Markdown guides. Their AI companions instantly became personal game-building wizards.

Soon the hands flew up: “Teacher, how do I do this?” My reply never changed: “Ask the wizard.” The agent parsed their intent, wrote code, and Live Server pushed changes straight to the browser. Total absorption.

One highlight: I wiped an agent’s short-term memory and said, “We’re handing the project to a new helper. Summarize what you’ve done and craft a message for the next AI.” Panic! “But what about our work? Can’t we keep our wizard?” Adorable proof that humans bond with AI personas.

Next, I had them ask a second agent to critique their game. Grumbles: “What does this new guy know?” Exactly the discomfort I wanted—they were learning to welcome critical feedback.

Once the framework was humming, my role shrank to wiring a pipeline that publishes their games to GitHub Pages, play-testing, and cheering them on. They were so engaged I had to pry them away from the keyboards.

Prepping all this while juggling my full-time job wasn’t easy, yet the payoff was huge. Frameworks that teach the art of questioning and collaborating with multiple AI personas feel like essential life skills for the age we’re entering.

telegram-cloud-photo-size-5-6086847233103348485-y.jpg

도경: https://ryoo8099.github.io/saving-my-own-cat/ 준혁: https://jun-hero.github.io/jun/hero-tower-defense.html 민교: https://guihongmei.github.io/piaominqiao/parkmingyo-game.html 성익: https://ksbom1.github.io/Monkey-fire-game/ 지아: https://jia1207.github.io/jjikin/ 민주: https://song2505.github.io/angel-cave-adventure/ 아리: https://masterclark-sanskids.github.io/tiger-shooter-game/ 요한: https://skim21.github.io/john/ 한울: https://hanul1.github.io/space-shooter-game/