




I built an open-source pipeline that turns PDFs into Anki decks — and makes the AI check its own cards against the source (MIT, free)
Like many here I got frustrated with "ChatGPT, make me flashcards" — you get 50 cards, half are vague, some are invented, and none follow good card design.
So I built anki-card-forge: you drop a PDF into a folder, write one sentence in the chat, and Claude reads the source, writes the cards following evidence-based rules (atomicity, active retrieval, no hint leaks), and then has to pass its own quality gate:
- Grounding check — every answer is verified against the source text, so hallucinated "facts" surface before you learn them
- Visual self-review — cards are rendered as PNGs (light + night mode) and inspected, incl. image-occlusion mask placement
- Real-engine validation — every deck is imported and rendered with Anki's actual backend before it's done
It also does image occlusion from figures in your PDFs, can push finished decks straight into Anki via AnkiConnect, and can restructure already-learned decks without losing scheduling (GUID-preserving rebuilds).
The GIF shows a deck built from the Wikipedia article on electric current with one prompt — the walkthrough with screenshots is in the README.
Costs & license: the tool is free and MIT-licensed. It runs on Claude Code (Anthropic's coding agent), which is a paid product — so this is for people who already have that. No extra API key, nothing leaves your machine except what Claude reads.
Disclosure per rule 6: built largely with Claude Code itself, and the card authoring is AI by design — the point of the project is forcing quality checks around it.
GitHub: https://github.com/FrostySL/anki-card-forge
Would love feedback — especially on the card-quality rules, and what checks you'd add.