掲示板 Forums - Stories generated from a list of vocabulary

I've been learning Japanese for the past two months, and I'm completely absorbed. I'm trying to use various tools (including renshuu as my daily driver), however, I'm having a hard time finding texts that are more than just single sentences, but don't stray too far from the vocabulary I've already mastered. While browsing Renshuu's settings, I noticed that I can export a list of words I've learned (tools -> advanced search), so I thought, maybe it's worth using AI to create short stories based on this list? I wrote a prompt that you can configure, you can find it below. It's certainly not perfect - I'm no expert - but if it's useful to anyone (or someone is willing to improve it), I'd be very happy. :) I tested it on the free version of ChatGPT, but it's probably suitable for other LLMs as well, so pick whichever works for you.

PS. As I'm a newbie, possibly this may be a well-known topic and every learner does it (and I simply missed it in my ignorance), but it was "new" to me, so I decided to share. I hope encouraging the use of AI doesn't violate renshuu's rules.

My prompt (don't forget to attach vocabulary list export to your prompt!):

Japanese Reading Practice from My Known Vocabulary

You are my Japanese language learning assistant.

I am learning Japanese and have attached an export of the vocabulary I have already studied in Renshuu. Use this file as the primary source of my known vocabulary.

Configuration

Adjust the following parameters before generating the stories:

* Number of stories: 3

* Story length: short, approximately 100–150 Japanese words per story

* Japanese level: beginner/intermediate — use simple grammar appropriate for JLPT N5

* Story topics: everyday life, school, work, travel, hobbies, or other realistic situations

* Difficulty priority: vocabulary from my attached list should be prioritized over introducing new vocabulary

* Story variety: make the stories different from one another in both topic and vocabulary where reasonably possible

Vocabulary rules

1. Use vocabulary from the attached Renshuu export as much as possible.

2. Do not intentionally introduce vocabulary that is not present in the attached list.

3. If a new word is genuinely necessary to make a natural and understandable story, you may use it, but keep the number of new words as low as possible.

4. Every new word must be listed before the story with:

* the Japanese word,

* its furigana,

* its English translation.

5. Do not count grammatical inflections or conjugated forms of words from my known vocabulary as new vocabulary. For example, if I know 食べる, forms such as 食べた, 食べない, or 食べました should not be treated as new vocabulary.

6. Prefer natural Japanese over awkwardly forcing vocabulary into the story.

7. Do not use furigana inside the story itself.

Required format for EACH story

### Story [number]: [Title]

#### 1. New vocabulary

List only the words that are genuinely new to me and were not present in the attached vocabulary list.

Use this format:

* 新しい単語 — furigana — English translation

If there are no new words, write:

> No new vocabulary.

#### 2. Known vocabulary used in the story

List all words containing kanji that were used in the story and that come from my known vocabulary.

For each word, provide:

* The exact form used in the story, including any conjugation or inflection;

* its furigana.

Do not provide English translations in this section.

Example:

* 学校 — がっこう

* 行きました — いきました

* 先生 — せんせい

Important: Make sure every Japanese word containing kanji that appears in the story is included in this list with the correct furigana for the exact form used.

Do not include words written entirely in hiragana, katakana, or other forms without kanji.

#### 3. Japanese story

Write the complete story in Japanese.

* Do not use furigana.

* Do not add vocabulary notes inside the story.

* Use natural Japanese.

* Keep the grammar within the configured level.

* Prefer vocabulary from my attached list.

#### 4. English translation

Provide a natural English translation of the entire story.

Do not translate sentence-by-sentence with additional grammatical explanations unless necessary for clarity.

## Important accuracy checks

Before giving me the final answer, verify that:

1. Every word marked as "new vocabulary" is actually absent from my attached vocabulary list.

2. No unnecessary new vocabulary has been introduced.

3. Every kanji-containing vocabulary item used in the story appears in the "Known vocabulary used in the story" section with correct furigana.

4. The furigana corresponds exactly to the form used in the story.

5. The story itself contains no furigana.

6. The Japanese grammar matches the configured level.

7. The stories are natural and understandable rather than artificially constructed just to use vocabulary from the list.

If the attached vocabulary file contains multiple forms, readings, meanings, or other metadata, use the Japanese vocabulary and readings from the file to determine whether a word is already known.

Do not assume that a word is new merely because its exact conjugated form is not present in the file.

---

Output structure

Repeat the following structure for each story:

# Story 1: [Title]

## 1. New vocabulary

...

## 2. Known vocabulary used in the story

...

## 3. Japanese

story

...

## 4. English translation

...

Then continue with Story 2 and Story 3 using exactly the same structure.

If you want to do this, you’ll probably be better off with an LLM trained specifically for Japanese translation, like sakana.ai. (Not an endorsement)

Guys, thanks a lot for you answers, it really got me thinking! It looks I just need solid N5 reading materials + Japanese dictionary from time to time. I have a feeling that my idea has some potential, but I won't base my learning habits just on a rookies hunch, so I'll explore human-made materials first. :)

​I put a lot of effort to improve my learning methods, and AI helps me a lot with it, but I'm still looking for new sources of proper texts. I already use Shinobi and Crystal Hunters (which are both great but not infinite), I've tried to use Tadoku books but they're sometimes too confusing when I try to read them using "Tadoku rules", and now I'm thinking about simply printing transcripts from Slow Japanese with Yuki videos (which are currently my main learning tool next to renshuu) and using them for reading exercises.

​They can be good at this, but they can also just make stuff. It's impossible to know which it did if you don't already know Japanese. It's also worth keeping in mind that most LLMs were trained mostly on English. Japanese represents a small fraction of their training data, so expect them to be worse at Japanese.

As an example of the limitations of LLMs, ask it why 今朝毎朝 is funny. I bet it'll say something about oxymorons, but in reality it just sounds like an inappropriate English phrase. This is a particularly hard example for LLMs, but it does highlight the issues with relying on them.

I do not think that presuming a "They learned before LLMs, therefore they cannot/will not see the benefits of LLMs" is not a fair take.(Obviously) I also work in software engineering, and comparing that to language learning is apples and oranges. While it naturally changes with the scope, scale, and complexity of a request, it is more common than not to be able to ask an LLM for some code, then have a method independent from the LLM to almost immediately verify that that code works. This is because, (again, speaking generally), you (the LLM user) are competent in the expected output. If you ask an LLM to make a tool that crops images, you do not need to know the code to be able to look at the outputted tool and say "yes - that is cropping it", or "no, it's not."

Using it for language learning is a whole different ballgame. There is no sense of actual confidence in the LLM output (unless you are using it more as a search engine to reach pages that have the info you're looking for), but if you are asking it to explain a nuance or something, you (presumably) do not have the knowledge to be able to judge the accuracy of that explanation. So if and when you get bad info, that's going to go right in there with all the correct info, waiting minutes/days/weeks/years for some other input to correct it, then the required time and energy to overwrite that misconception.

I still do believe that people should use what works for them, and renshuu (as a community) does not block out certain types/styles of learning - we very much embrace a "you do you" thing.

However, I don't think the comparison between these two fields is a valid one, and for those that do choose to use LLMs for language learning, they should have abundant awareness of the dangers and shortcomings (just like with any learning method). A similarly styled argument could be made for learning via exposure to anime. Not that the benefits and drawbacks are the same, but that it is a method that has both good points and bad ones.

The best way to utilize any tool is to understand these benefits and pitfalls, and then use it with both of those in mind.

​If that's the case, it's a recent change. I test this example occassionally just to make sure they still consistently fail. Just to be clear about the prompt, I always ask something like "why is 今朝毎朝 funny?". If you give them enough context, they can put it together. But at that point, you're explaining the whole pun yourself.

I'll have to test again. It's possible they're just looking it up now instead of generating the answer from scratch.

​Yes, I'm still looking for new input, please share it. :)