Questions before installing

It is for people who want to learn from material they choose themselves and keep reviewing it over months or years. You pick words from books, subtitles, lessons or conversations, and kamishibai turns them into cards. It does not replace a course: it does not choose your topics or the order in which you study them.

Anki handles the reviews. Its schedule uses spaced repetition: cards you remember return at longer intervals, while forgotten ones come back sooner. You set the number of new cards and daily limits in Anki.

I believe the new generation of AI agents is making it possible to create software of our own as easily as we once wrote blogs: sharing our experience through approaches other people can easily put into practice. Kamishibai is my way of putting that idea to the test. Above all, it is a tool I use myself every day and improve based on that experience.

A mangaka personally brings a story, drawings and characters together into a coherent work. That kind of authorship appeals to me. It also explains the visual style of kamishibai: I like a project to have a point of view you can recognise in both its underlying idea and its visual design.

In the main workflow, you add a list of words or expressions you want to learn. You do not have to put them in dictionary form: kamishibai tries to recognise what you entered and offers meanings to choose from. Each selected meaning gets a separate card with a phrase.

If a phrase does not fit, describe what you would like changed and generate it again. If you already have complete card text, you can supply it as cards JSON. There is a separate answer about that below.

Kamishibai prepares the material: phrases, explanations, audio and illustrations. It packages them into an .apkg file, a deck you can open in Anki.

Anki handles the reviews: it shows the cards, records your ratings and schedules the next review.

Kamishibai also creates a PDF for paper cards. You can print it and cut out the cards if you prefer to study away from a screen.

No. After installation, you run kamishibai in a terminal, a window with a text interface. The app then guides you through adding words, choosing meanings, checking the cards and getting your deck. You do not need to write code.

Separate commands that return JSON are available for automation. Scripts and AI agents can use them, but you do not need them for everyday use.

You choose the words and meanings. AI models create the material according to the rules built into kamishibai. Those rules define the card structure and the requirements for its phrase, explanation, audio and illustration.

Before generating cards, you can choose the difficulty and phrase type: statements, questions, dialogues or a mixture. You can check the result and ask for another version with instructions about what to change.

There are 22 supported languages: English, Chinese, Spanish, Japanese, French, German, Korean, Russian, Italian, Portuguese, Hindi, Arabic, Turkish, Polish, Ukrainian, Indonesian, Vietnamese, Thai, Greek, Hebrew, Dutch and Czech.

Each deck uses one pair: the language you are learning and the language for translations and explanations. You can combine any two languages on the list. The quality depends on how well the models work with that pair.

If you need another language, open a request on GitHub and we can see what is possible.

To create cards, you need:

• A computer running macOS, Linux or Windows.
• An internet connection: text, audio and illustrations are generated through the Gemini API.
• Your own Gemini API key. Create a key in Google AI Studio →

To review the cards, install Anki →

If the selected language needs local text recognition to check the pictures, the app will also download OCR models the first time it creates illustrations.

Prebuilt versions are available for macOS, Linux and Windows.

• macOS and Linux: install with Homebrew or the installation script.
• Windows: run the installation script in Git Bash, MSYS or Cygwin.

You can find the commands and script link in the installation section.

Kamishibai is free and open source. Gemini requests made with your API key are charged separately.

As a guide, expect around $15 for 100 cards with text, audio and illustrations. This is an estimate: the total depends on the amount of material, how often you regenerate it and model pricing. The app already has its models configured, so you do not need to choose them to get started.

At four new cards a day, a hundred-card deck gives you 25 days of new material. You will keep reviewing those phrases afterwards. That is why I choose words I actually want to use and put card quality ahead of the lowest possible generation cost.

Your work is saved on your computer: settings, sessions, cache and finished files. Kamishibai has no server of its own for creating cards.

To generate and check material, the app sends words, phrases and illustrations being reviewed to Gemini using your API key. If you sync through Anki, the finished deck and media also go to AnkiWeb.

A saved key is kept in a local settings file, not in the system password store. Do not share that file or put it in a public repository.

Yes, mistakes happen. In my experience, though, the current generation of models already produces fairly consistent results, and I put a lot of work into the quality of the finished cards.

Audio is currently the most reliable part and sounds close to natural speech. Models are less consistent at matching phrase difficulty: they aim for your chosen level but sometimes produce a phrase that is easier or harder.

Illustrations are the most prone to errors: AI can still draw a person with a third arm. That is why the app uses separate AI models to assess the finished pictures. If an image fails the check, the app generates it again.

Automatic checks cannot guarantee that everything is correct. If a phrase seems wrong, check it with a dictionary or a teacher and generate a new card if needed.

An AI agent, such as Claude Code, can use kamishibai. You describe the task, and the agent carries it out through the app's commands. For example, you can ask it to:

• Pick words on a topic or from a text you provide.
• Choose the relevant meanings and create cards.
• Check the results and regenerate unsuitable cards before importing them into Anki.

This works through the JSON CLI: commands that return JSON data without prompting for input in the terminal. An agent can start generation in the background and run several sessions at once.

Give the agent the link to llms.txt. It explains how to install or update kamishibai and get the built-in instructions for the installed version. Using the agent itself and generating extra cards can increase the cost.

When you want to keep existing wording, such as sentences from a lesson or short dialogues. Instead of a word list, you supply the card content in a cards JSON file, a text file with a defined structure.

The file must include every required card field, not just a phrase. You can prepare it yourself or ask an AI agent to do it. The wording in the file is kept as written. Kamishibai adds audio and illustrations, then builds an Anki deck in APKG format and a printable PDF.

Schema, required fields and examples in cards-json.md →

Yes. Kamishibai creates the cards on your computer, and you can review them in an Anki mobile app.

• Open the finished .apkg file in Anki on your computer. If double-clicking does not start the import, choose File → Import.
• Sync your collection through AnkiWeb.
• Sign in to the same account in the mobile app and sync to download the cards, pictures and audio.

To study away from a screen, you can print the PDF, cut out the cards and fold them. Paper cards have no audio or automatic review schedule.

Models from the Gemini family generate the text, audio and illustrations. Kamishibai gives them instructions, checks parts of the output and assembles the material into cards.

The audio is synthesised by a model, not recorded by a narrator. Check it along with the rest of the content before you start studying.

If you cannot recall the meaning of a phrase or its key word, choose “Again”. Anki will bring the card back sooner. This applies to a new card just as much as one you have known for months.

How strictly you rate yourself depends on what you are practising. At first, I allow a small article or preposition mistake if I remember the meaning and key words. As the phrases become familiar, I start expecting accuracy in those details too.

If one card keeps giving you trouble, look at the phrase itself: is the situation clear, and is there too much unfamiliar material in it? Sometimes simplifying or replacing the example is more useful than trying to memorise it again as it stands.

There is no set timetable. It depends on your starting level, the phrases you choose, how regularly you review and whether you use the language outside Anki.

You could use your first 50 well-learned phrases as a point to check your progress: are familiar expressions easier to recall, do you recognise them in new texts, and can you use them in conversation? This is not a threshold that automatically gives you a “feel for the language”. Look at what you can do, not just your card count.