guanjiawei

Posted on May 1 • Originally published at guanjiawei.ai

I Spent a Day and a Half Using AI to Make a Literacy Game for My Son

#ai #family #education #thoughts

Recently I gave a talk at a major conference. Before and after the event, I met quite a few people. Among the attendees, there were broadly two camps.

One camp spends every day immersed in open-source communities, building products and growing their own direction. Chat with them for a few minutes and you can feel a density of being on the front lines—their viewpoints grow on-site, carrying heat and contagious energy. The other camp carries a different scent: boss-assigned tasks, checkbox logic. They talk about "I built a thing, it can do this and that"—a very standard past-tense software narrative. Right now, AI is least lacking in "what it can do"; what it's most lacking is "whether it can do it well." The latter camp usually runs out of things to say right there.

There's already a sizable gap between these two groups. The bigger chasm is still behind them; I'll come back to that. First, let me talk about what happened at my home these past two days.

"Why Are You Always Staring at the Computer?"

Work intensity has been higher than usual these past few months. Not the busy-ness of office hours, but the busy-ness that comes from internal drive. The boundaries of agents shift every day; exploring them is addictive, and I can't stop. I leave early and often get home past eleven, constantly pushing family matters to the back of the queue.

A few days ago my son suddenly ran over and asked me: Why are you always staring at the computer?

I didn't brush him off. I had him sit down and explained, in words he could understand, what artificial intelligence is. He stared at the cursor moving on its own across my screen and asked: How is the computer typing by itself? I said: I told it what I want, and it's helping me work.

He was still a bit confused. I said, let me make something for you right now.

At that time he was obsessed with PAW Patrol. I opened a coding agent and dictated to it: I want a PAW Patrol Mighty Pups battle game, landscape screen, where you can pick a pup to fight. Then I had him go play something else, and said I'd call him back in five minutes.

Five minutes later the game was actually running. Six pups to choose from, light punch and heavy punch, crouch and jump—all working. He played with an incredulous look on his face and then told me: Huh, seems like Mighty Pups' special move didn't come out. I said, let's fix that. Five minutes later, the special move was there too. The effects were crude, but enough for him.

He wanted more. I said: When you get a bit older, Dad will make you something even better.

A Kid Who Couldn't Get the Little Red Flower

Probably that impromptu game made him feel like Dad was finally participating in his world for once. He asked if I could teach him to read a clock—the kind with hands (that thing is actually very unfriendly to kids; that's another story). In the middle of teaching he lost interest, and instead asked if I could teach him to recognize characters.

I was a bit surprised. He quietly said that at school you can get a little red flower for character recognition, and he often couldn't get one.

I understood that sense of frustration immediately. I asked him why he didn't tell Mom. He couldn't say; he just sat there sulking.

In that moment I made a decision: take the same impromptu-game approach and apply it to "teaching him characters." Make a literacy game just for him, using his favorite IP, where adults can play along.

Don't Write Code Yet—Let the Agent Do Its Homework First

That night when I got home, the first thing I did after opening the coding agent was not to ask it to write code, but to have it do background research.

I threw three questions at it:

What are the recent pedagogical and psychological conclusions on Chinese character learning for preschool children?
What memory-related learning theories and methods are suitable for preschoolers?
What's the deal with the PAW Patrol IP, and what is its narrative structure?

It spent about an hour on research. When it threw back its summary, I realized several of my default assumptions were wrong.

The preschool focus isn't pinyin. Starting in the fall of 2016, the Ministry-compiled first-grade Chinese textbooks swapped "a o e" for "sky, earth, human"—characters first, pinyin later, putting pinyin back in its place as "a tool to aid character recognition." This wasn't a snap decision. In 2002, Guangdong Provincial Women and Children's Hospital ran a controlled experiment in kindergarten senior classes: the Chinese-character class outperformed the pinyin class in both scores and speed. In 2011, Beijing Normal University's School of Psychology followed 176 first-graders for a full year and found that children who consistently read pinyin-annotated materials actually had lower reading self-efficacy than those who read unannotated materials. What preschoolers really need to focus on are pictographic characters, letting kids move from "image" to Chinese character, not from symbol to character.

Don't be greedy. The Ministry of Education's 2012 Guidelines for Learning and Development of Children Aged 3-6 explicitly opposes early mechanical character memorization and intensive training. The time it takes a kindergartener to learn one character is roughly three to four times what it takes an elementary student to learn the same character. Method-wise, the three-bucket method (a simplified version of the Leitner system, a spaced-repetition method proposed by German science journalist Sebastian Leitner in the 1970s) is well-suited for this age. Bucket 1: review daily; Bucket 2: every three days; Bucket 3: every seven days. New characters go into Bucket 1; move back on correct answers, drop to Bucket 1 on mistakes. Keep daily new characters to two or three—much steadier than cramming ten at once.

PAW Patrol itself follows a fixed formula. The rhythm of nearly every episode is identical: a resident gets in trouble → call Ryder → assemble the pups → pick the right rescue pup → head out → rescue process → celebration wrap-up. This highly repetitive structure is actually very suitable for level design. Each level is a complete "alarm, dispatch, celebration" cycle. The rhythm is familiar, frustration is low, and kids won't get blocked by unfamiliar new logic.

After clarifying these three things, I started designing the game with it. For this segment I had it use the "superpower" skill to do the design. It would frequently pause to confirm with me: "Does this mechanic feel OK?" "Should we cut this step?" After a few back-and-forth rounds, a runnable framework emerged.

Making Assets Is Way More Exhausting Than Writing Code

A framework is one thing; assets are another. Subtitles, images, sound effects, character voice-overs, level backgrounds—you need everything. I originally wanted to record the voices myself, but gave up after a few tries.

I had another agent dedicated to making assets. It built the entire pipeline itself:

Find original episodes on YouTube and extract the audio.
Run ASR locally, split the audio into 2-second clips, and extract the dialogue for each clip.
Use a vision model to extract frames from each clip, identifying which pup and which scene appears on screen.
Run TTS locally to generate in-game character prompt voices.
Generate visual assets using Gemini's Nano Banana (at the time, GPT's new image model hadn't come out yet).

At first it wanted me to manually tell it which pup was in each frame. I couldn't recognize every PAW Patrol character and got confused. It revised its own approach: extract text via ASR first, then use the vision model to align on-screen content with the dialogue, filtering automatically. I reviewed this pipeline, saw it was feasible, and let it run.

I tried the first version and found a pile of bugs; the interactions were so complex that even I didn't want to touch it. I had it cut a batch of features and simplify to the point where a child could pick it up alone. The next day another version ran, still with many issues.

At this point I did something that, in retrospect, was pretty crucial: I had the agent run a round of end-to-end testing itself before waiting for me to test. After running through, it listed a whole stack of issues on its own, and while it was at it, asked, "Should we add an art director role to unify the visual style?" After adding that, the visuals immediately became cleaner and the asset consistency improved a lot.

The whole process took a day and a half. By the second night, it was something playable.

The First Time I Let the Two of Them Play

I got home after 11 p.m. that night and asked my wife if she wanted to try it first. It was too late, so we didn't. It got pushed back a few days.

Yesterday afternoon she suddenly remembered and asked if they could play. I wasn't home, so I used Tailscale to tunnel from her phone into my computer and started the game remotely.

She ran through it herself and texted me back: Better than I expected. One design I particularly liked in this version is that the child looks at one screen while the parent, acting as "commander," looks at another screen on their phone. The parent receives rhythm prompts telling them when to encourage and when to help. She was completely on board with this split-screen setup.

Then she played with our son. After finishing, he refused to leave and said he wanted to play again tomorrow. Perfect—according to the theory, three characters a day is enough; more than that is greed.

The Look on Her Face That Moment

The most unexpected part of this whole thing wasn't my son's reaction; it was my wife's.

I had talked to her about AI many times before. About agents, about coding, about what was coming next. She listened, but didn't feel it—much like my son the first time he heard me talk about artificial intelligence. When a concept doesn't land on something she cares about, no matter how much you talk, it's someone else's business.

This time was different. Their child's education is the heaviest thing in her heart. Seeing something made in a day and a half actually get her son to sit down and play, her reaction wasn't "that's amazing"—it was "I can use this."

That night she said something to me: Then can we make one for our friends' kids too?

I said yes.

Three Types of People, Two Chasms

Back to the conference I mentioned at the beginning. Inside the conference hall there were two camps; outside in the wider world, there are actually three types of people.

The first type has no real awareness at all. They know AI is hot, they've skimmed plenty of articles, but they feel it's far away. They don't get their hands dirty, and don't feel they need to.

The second type has gotten their hands dirty, but they're doing tasks someone else assigned. The boss says the company needs to build something like this, so they build it, check the box, and it's over.

The third type uses it on something they truly care about. In the process they feel a boundary get pushed open, and then they can't stop.

Between the two chasms, the second is larger than the first. The first is merely the distance between "having heard about it" and "having tried it." The second is the distance between "checking a box" and "being ignited." The checkbox people finish one round and clock out; the ignited people go home still thinking about the next step.

In the past, when people said "using AI," it was easy to interpret as a skill: knowing a few more prompts, being familiar with a few more tools. Over the past couple of years I've increasingly felt that's not the key. The key is: put it into that thing you lie in bed thinking about every night.

The heaviest thing in my wife's heart is our child's education, so that's what went into the PAW Patrol literacy game. I work on AI Infra; the heaviest thing in my heart is hardware efficiency and TCO, so that's what goes into the endless experiments I run every day. It doesn't need to be something far away—just something you truly care about. At that point it will quickly show you what it can do. And the way it tells you isn't through a few articles; it's through you personally making something you previously wouldn't have dared to imagine.

Products like BaiCiZhan used to require a team, a company, building something for everyone to use. Now it's a dad spending a day and a half making a personalized literacy game for his own son. These aren't the same kind of thing; they're two different worlds.

First find that thing you truly care about. The rest of the path will grow on its own.

References

Originally published at https://guanjiawei.ai/en/blog/paw-patrol-literacy-game

DEV Community