How My Anki Flashcard Generator Works

My wife and I love France and wanted to get better at the language, so we started lessons with Bill Savage. (He’s excellent, by the way.)

For memorization, I love Anki flashcards, but they’re a pain to make, so I vibe-coded a CLI tool that automates the most annoying parts. Here’s how it works:

1. During lessons, Bill sends WhatsApp photos of the whiteboard notes.

2. After the lesson, I drop those photos into a Claude Project that’s set up to spit out a markdown file with one term per line.

3. I run a CLI script that loops through each term and calls the Anthropic API. Claude returns structured JSON for each word: gender, pronunciation, French/English explanations, and 3–5 example sentences (line-matched across both languages).

4. The script validates every flashcard for missing fields, wrong example counts, and malformed data before showing me anything.

5. Flashcards print to the terminal so I can scan them. Anything broken gets flagged.

6. After I confirm the output, the script sends the flashcards to Anki via the AnkiConnect plugin. It auto-skips any duplicate terms.

7. Every batch gets logged to card_log.jsonl with note IDs. If a batch turns out bad, --rollback deletes the last import.

8. I hit Sync in Anki to push to AnkiWeb.

9. Study Anki flashcards!

The full stack is Python, Anthropic SDK, and AnkiConnect. Runs me about $1.40/month in API costs for ~20 flashcards/week. You could argue that creating flashcards manually is better for learning but in practice the manual effort caused me to 1) make fewer flashcards 2) the flashcards I did make were less useful (e.g. no pronunciation or example sentences).