The app draws a random Japanese word from a predefined list, and asks me to type the romaji (romanization) of the word.

For example, ホテル = ho-te-ru which means “hotel”. Simple, right?

Turns out I had to navigate waves of challenges before reaching my goal. Here is the story of me learning Japanese and agentic AI workflows simultaneously.

3-point summary, human crafted ;)

• GenAI helps experts soar while entrapping amateurs in pipe dreams.

• As soon as I ventured into areas beyond my tech expertise, AI hallucinations and my inability to diagnose errors began hindering the development process.

• Nonetheless, agentic AI sped up development by 3-5x. Time to a Minimal Viable Product stands: 8 hours.

  • Model chosen: gpt-oss-120b through Ollama cloud.

  • Yes, you can use Claude Code without Claude aka Anthropic’s proprietary LLM. Henceforth in this article, “Claude” is short for “Claude Code”.

Share

Subscribe now

Wave 1: choose a tech stack (easy, right?)

THE CATCH

I knew nothing about writing web apps before starting this project. As a data scientist I may have created the occasional Streamlit app, but that experience is too infrequent to provide a mental model for “what to do” in general.

THE SOLUTION

I set Claude to Planning Mode and asked it to consider web app frameworks that are 1) beginner-friendly; 2) suitable for local deployment. It settled on Flask and rejected Streamlit, because of Flask’s minimalist nature.

Personally I think it is OK delegate the choice of framework to Claude, because—

1. it is likely more knowledgeable than a complete novice like me;

2. the justifications are easily verifiable.

Then, whenever I ran into bugs, I first let Claude identify their origin—and then pushed myself to manually inspect and maybe fix them. At first I tried to stay entirely within Claude for my workflow, but it was unable to fix many runtime issues that show up visually .

Maybe I should have given Claude access to my screen; the most it can see are diagnostic and error messages on the terminal. However being the Luddite that I am, that is not going to happen in the foreseeable future.

Nonetheless, the knowledge distilled from manual bug fixing has helped me understand my app and Flask better.

THE OUTCOME

Flask app running locally on a given port, accessible through a browser

Decorated Python functions in main.py provide all functionalities. For example, Claude decided that sample() would handle POST requests to /check.

@app.route(”/check”, methods=[”POST”])
def sample()
    ...[Determine correctness of user input]...

The visual interface, on other hand, is provided by Claude-generated HTML+CSS, within which are buttons that trigger GET requests to different parts of the app e.g., /check, /reset.

Share

Subscribe now

Wave 2: hiccups in Japanese — the AI’s and mine

THE CATCH

By my own design, I am supposed to provide the app with a list of Japanese words and their “ground-truth” romaji or romanization. The app then checks user input against the romaji of each word, and indicates “correct” vs “wrong”.

In practice, 60% of the list was generated by an LLM during development. Automation, yay?

The problem is … I don’t know how to express certain common sounds in romaji. To make matters worse, the LLM I chose isn’t completely solid or consistent on the subject either.

For example — がっこう（学校、school）where っ is silent and indicates a brief pause. How do you indicate silence on a keyboard?

Another example— きょう（今日、today）where the small よ assimilates into the preceding き. Am I supposed to type them as separate kana or some amalgamated version?

THE SOLUTION

Turns out, non-GenAI expert systems contain the requisite knowledge. I just need to tease it out of them.

My computer’s built-in Japanese input method came to rescue. With some trial-and-error I identified the multitude of letter combinations that produce the desired words and sounds.

Google Translate served as an addition source of knowledge For example, it represents がっこう（学校、school）as gakko, where the silent tsu っ became the first of the two k’s.

It turns out to be a pattern: silent っ → double the non-consonant letter of the next kana; in this case ko こ. Once I can confidently type out similar words, I included them in my app.

THE OUTCOME

13 words in hiragana and katakana, respectively, giving a total of 26 words.

It is by no means scalable. I will find a way to teach an LLM to consistently notate romaji.

Wrapping up: model choice, good practices

After overcoming Waves 1 and 2, I have arrived at a Minimal Viable “Product”— of which a screen recording can be found at the top or right below.

The whole exercise is a novice’s attempt to learn both Japanese and agentic AI workflows. Let me end this post with lessons I gleaned from building the web app.

1. I used gpt-oss-120b through Ollama cloud, and I now wonder if that is the most cost-effective option.

   1. While cheap relative to my usage—the Free plan is enough for a couple hours of hacking each day—it did hallucinate a few times, claiming to have accomplished things whilst sitting on its ass, sprouting confident BS.

   2. Anthropic has just updated Claude models to achieve the same outcomes with less agentic iterations and therefore tokens. In a nutshell: replacing a linear agentic loop with simultaneous instances of code writing + tool use before injecting context into an LLM.

2. Regardless of LLM chosen, it helps to /clear context and restart the chat at every “phase” of development.

   1. The purpose is to avoid encumbering the LLM with irrelevant context. This is especially important if you have a somewhat “dumb” model.

   2. The break points could be the transition from planning to initial implementation; or it could be as simple as adding a new feature.

   3. My unscientific heuristic: clear context before it gets 50% full.

And finally, be patient! This is new technology, bugs and hiccups are bound to crop up eventually.

Focus on boosting output while retaining human-in-the-loop. That, is how we grow creative and productive without incurring brain rot.