海馬：AI 如何改變我學習程式語言的方式

1. In one sentence

A primary-school developer, using only the free tiers of Claude AI, Gemini, Grok and ChatGPT, compressed a learning curve that once took six months to produce an app into one month to produce a complete brand website.

2. The session

海馬 introduces himself as 海馬 of "海家帶你游程式", with interests in programming and video games, already shipping an app called "9歲記帳" and currently doing live streams of AI-assisted coding. He said today's talk would cover the four AIs he has been using, the brand website he built with vibe coding, a comparison between how he learned without AI and with it, plus his GitHub repository and his latest Claude AI prompt.

Building a brand site with four free-tier LLMs

He set out how he divides the work between AIs: Claude AI for more complex code; Gemini also for code, his first choice when Claude's quota runs out; Grok for simple code when ChatGPT and Gemini are both out of quota; and ChatGPT for debugging simple code. He stressed that all four are the free tiers, and with those four free-tier LLMs he built the "海家帶你游程式" brand website in a month.

He describes working top-down, cutting the problem into small pieces: first give the AI a screenshot of the design and ask it to lay it out; then add the headline text with a specified size; add buttons and icons, giving filenames for the AI to lay out; when a colour has to be exactly right, give the colour code (#80B5FF, say) rather than only an image for the AI to pick the colour from; add a line of copy at the bottom; and when the press coverage section's layout broke and images were cropped, he learned to describe the requirement with a technical term like "adaptive", which the AI understands better than "fill the screen". Finally, when the text all bunched up illegibly, he gave the AI two comparison screenshots so it could see exactly what differed between them, and that fixed it. He initially chose not to do responsive design (RWD, where the same site adjusts its desktop and mobile layouts to the screen size).

He said that at one point he had no idea the text he had added had been placed behind the background image, and looked more than ten times before finding the problem — which is why he later says "even using AI, you still have to learn some programming": if you have learned a little, the AI can fix things quickly; if you have learned nothing, you do not even know where your own things have been put, do not understand terms like responsive design or adaptive, and get stuck more easily while burning more free-tier quota.

From Apple's documentation to learning Python and SQL with AI

He shared three before-and-after comparisons from his learning:

• Apple's WatchApp development tutorial: the site says the tutorial takes about 25 minutes; knowing nothing at the time, he took a whole week. His method was Xcode on the left, the tutorial site top right, Google Translate bottom right; he found that translating the whole page gave very strange renderings of technical terms, so he only translated the parts he did not understand. With AI available at the time, he would have translated with ChatGPT, because ChatGPT can list the steps briefly and clearly in a way Google Translate cannot.
• Learning Python: he used the book *Mastering Python, second edition*, reading and typing out every example himself. He carved out time every morning after six, before school. One early morning he hit an error, asked his father, who was still in bed, and was told to type it out ten times first — and all ten were still wrong, because it turned out the book itself had a misprint (he has been collecting an errata list ever since). With AI at the time, he would have screenshotted the error and the AI would have pointed out a misplaced comma and fixed it.
• Learning SQL: he watched a three-hour beginner SQL database tutorial video and afterwards designed his own expense-tracking table, with columns for amount, payment method (EasyCard, foreign currency) and category (food and fun, shopping, and so on). With AI at the time, he would have explained he was building an expense-tracking app, handed it the table, and had it generate the corresponding features directly.

Traditional methods versus learning with AI

He put together a time comparison: for "9歲記帳", learning Python took three months, hand-building the interface two months, and developing features with AI one month — six months in total; for the "海家帶你游程式" brand website, learning web development and building the site were both entirely AI-assisted, about a month in total. He also listed pros and cons: traditional methods need rest, AI does not; traditional methods mean a lot of Googling, asking AI is direct and fast; traditional methods involve a real person helping, AI is virtual. His conclusion is to choose AI, because it is fast, available around the clock, and you do not have to worry about waking your father while he is resting.

But he also stressed that having AI does not mean not needing his father: the app submission process involves a pile of settings where AI actually makes it more awkward; Xcode's basic interface operations are too hard for him to understand from a text description; and when a GitHub push failed once, it was his father who told him it was because certain things are no longer supported — a situation AI is less able to help with. His conclusion: his father is not going to be replaced by AI just yet. Checking press coverage afterwards, he was 9 when he gave this talk and had turned 10 by the time of later interviews, which is why the age varies between reports.

Testing Claude 4 live

He mentioned Claude 4 had been released only days earlier, and he could not wait to test it on "9歲記帳" — having it read its own code and add features. He said Claude 4 was too busy to get into at first and he only got a slot after his shower. He demonstrated a newly added theme setting live: you can pick a background colour with a live preview beside the options, and he demonstrated green, then a blue-green gradient, settling on a dark blue gradient. He also showed a voice expense-entry feature: before he had even finished speaking, the system had determined the amount was one hundred and fifty and the item was coffee. He said he built five features in two hours in total, using only a dozen or so prompts that the AI actually acted on, because most of the time was spent waiting for it to run and clicking to grant permissions.

3. Figures and cases

• Built the "海家帶你游程式" brand website within a month using four free-tier LLMs (Claude AI, Gemini, Grok, ChatGPT).
• The "9歲記帳" app: three months learning Python, two months hand-building the interface, one month developing features with AI — six months total; the brand website: learning web development and building it, all with AI, about one month.
• Learning Python from a book, he typed the same code ten times and it was still wrong, because the book itself had a misprint.
• Testing Claude 4 live: five features in two hours, using only a dozen or so prompts.

4. Lines worth keeping

• "If you haven't learned anything, the odds of going round in circles are higher."
• "As long as it gets me to my goal, it doesn't much matter whose help it is."
• "AI is useful, so I choose AI."
• "Anyway, Dad isn't going to be replaced by AI just yet."

5. Tools and terms mentioned

Claude AI, Claude 4, Gemini, Grok, ChatGPT, Xcode, Swift, Python, SQL, Apple Watch (watchOS), GitHub, responsive design (RWD), adaptive layout.

6. Wider observations

The detail that stands out most is that he learned to communicate with AI in precise vocabulary — colour codes, the word "adaptive" — rather than vaguely saying "fill the screen". That is the same thing Vivi described from the other direction when she judged her own visual sense poor and learned to cross-use different LLMs to compensate: the more precisely you communicate with AI, the better what it produces. Reading the three "non-engineers coding with AI" sessions together, you find a salesperson, an NGO officer and a primary school student all describing the same set of principles — cut the problem small, say it clearly, remove the ambiguity — arriving at them by different routes from their own circumstances. 海馬's advantage is that he is not afraid to start over and not easily fooled by AI (he says outright his father has not been replaced), and that pragmatic sense of proportion is more persuasive coming from a primary school student than it would be from an adult.

7. Sources