I’d like to introduce Daily AI Reading Reinforcement (DAIRR), an open-source Anki add-on that turns the cards studied today into personalized reading material.
Why I built it
Flashcards are effective for retaining individual facts and vocabulary, but recognizing a word on a card is different from understanding it inside a paragraph.
DAIRR is intended to connect those two stages. It collects the cards you studied during the current Anki day and uses an OpenAI-compatible model to generate a coherent article containing the selected terms.
DAIRR finds decks with study activity during the current Anki day.
You select a deck and choose the relevant card fields.
You can include all candidate cards or filter for newly learned, forgotten, or vague cards.
A prompt preset controls the article language, reader’s native language, difficulty, target length, and additional instructions.
DAIRR generates the article together with review notes for the source terms.
The result can be read inside Anki, exported as Markdown/HTML, or saved back into Anki as a suspended reading card.
Reading features
horizontal and Japanese-style vertical reading;
paragraph-level translation toggles;
separate article and review-note columns;
Chinese, English, and Japanese interface translations;
multiple visual themes; and
saved article history with deck/date filtering and a 52-week activity heatmap.
Provider configuration
The add-on works with OpenAI-compatible chat-completions APIs.
You can save multiple API profiles, choose models, fetch the provider’s model list when supported, and test the current configuration from the interface.
API keys remain in the local Anki add-on configuration. DAIRR does not provide or proxy an AI service, so users need to bring their own compatible API key.
It sounds interesting, but I’d rather it produced the additional reading material from a real source, rather than making it up. AI quite often gets things wrong or confused; your solution may end up reinforcing an incorrect understanding of terms, rather than supporting the good learning that you intend.
AI was not initially intended to use truth as a key criterion. I have seen numerous examples, where it has made things up, crafting together a whole lot of nonsense that may, at first glance, appear to be correct and on topic. It can also produce text with lots of fluff: i.e. text that uses the terms that you seek, but that text effectively says nothing. In the aforementioned two cases, these results will only impede learning, not support it.
Thanks for the thoughtful feedback. I agree that AI-generated material can sometimes be incorrect, vague, or misleading, particularly when the cards contain specialist terminology.
DAIRR was primarily conceived as a way to reinforce everyday foreign-language vocabulary by placing words due for review into a short, readable context. Using existing source material would improve provenance, but it would also be difficult to find a short real-world passage that naturally contains the particular combination of words a learner is reviewing that day. That would be a somewhat different workflow from the current purpose of the project.
The current version does allow users to submit a meaning field together with each term. The model can then use those supplied meanings as the basis for the article instead of relying entirely on its own interpretation. This does not eliminate the risk of errors, but it should reduce it. I’m also planning to strengthen the default prompt so that the model follows the supplied meanings more closely, and may add an advanced option for users to provide their own generation instructions.
Another possible improvement is to use the day’s Anki review results—Again, Hard, Good, and Easy—to assign a difficulty or priority score to each card. Users could then set a threshold so that only the more difficult cards are included in the generated passage. This would reduce the number of terms the model has to combine, making the resulting text more focused and reducing the pressure to force unrelated terms into a single article.
So I think the concern is valid, especially for professional subjects. For now, the project is best understood as a vocabulary reinforcement tool rather than an authoritative source of specialist knowledge, but your feedback has highlighted an important limitation that I should make clearer and continue working on.
Hi everyone — I wanted to share a major update to DAIRR.
DAIRR originally focused on turning the cards reviewed during the current Anki day into a contextual reading article. I’ve now expanded it into a broader reading and translation-practice workflow.
One motivation for this update was to make a back-translation-style exercise easier to do inside the same tool: read a translation, reconstruct the target-language text, get feedback, revise, and try again.
What’s new
Translation Studio
Practice from an existing DAIRR article by hiding the original text and translating segment by segment
Practice from pasted text, so the workflow is no longer limited to DAIRR-generated articles
Automatic segmentation, with manual editing/splitting/merging/reordering available before practice
AI review
Feedback is separated into multiple dimensions:
meaning
omissions/additions
grammar
vocabulary
naturalness
register/style
The review is meant to help learners improve accuracy and naturalness, not enforce word-for-word matching
DAIRR can still review a translation even when no reference translation exists
Configurable reinforcement scoring
Article generation now uses a more transparent and configurable scoring engine
Signals can include answer buttons, repeated same-day reviews, recovery after failures, FSRS-related values when available, duplicate/sibling cards, and recent use history
Signals are shown as applied, disabled, or unavailable instead of having missing data guessed or fabricated
Manual target planning
Cards can be marked as Required, Preferred, Optional, or Excluded before generation
I’m attaching a screenshot of the new Translation Studio reviewing a Japanese translation from pasted text.
