Claude AI Flashcard Auditor

An AI-powered quality checker for your Anki cards. Uses Claude (Anthropic's AI, a ChatGPT alternative) to spot badly designed flashcards during review, suggest fixes, and learn from your feedback over time.

Good flashcards follow specific design principles: atomic (one idea per card), minimal, non-circular, properly formatted. Breaking these rules costs you review time and retention. This add-on audits each card against those principles and fixes them in one click.

Features

• Dockable side panel next to the reviewer. Audit the current card with Ctrl+Shift+A or the Check Card button.
• Verdicts in seconds. Claude returns one of three results: ✓ Optimal (nothing to do), ✗ Delete — [reason] (card is obvious, circular, or not worth learning), or a one-line diagnosis plus a rewritten Front and Back you can apply with one click.
• Apply Fix edits the note in place. A before/after diff dialog lets you confirm before saving.
• Smart scheduling reset. When a fix materially changes the prompt or answer, the add-on recommends resetting the card's review history — because your past reviews were measuring different content. Minor edits leave scheduling alone.
• Editable system prompt. Tailor the audit rules to your subject (medical, language, law, CS, etc.) via a proper text editor in Settings — not raw JSON.
• Adaptive learning. Disagree with a verdict? Click 👎 and categorise why. After a few disagreements the add-on automatically sends them back to Claude to generate a "learned rules" addendum appended to every future audit. The auditor gets more aligned with your standards over time.
• Credit protection. Warns you before re-auditing a card whose text, model, and prompt are all unchanged, so you don't pay twice for the same verdict.
• Custom hotkey, custom model. Point it at any Claude model ID via Settings.

What it catches

• Non-atomic cards bundling two recall targets (definition + example, mechanism + effect)
• Circular backs that just restate the front in different words
• List cards where the count asked doesn't match the count given
• Redundant or overlapping bullet points on the back
• "Why/how" questions answered with a "what" definition
• Two-way definition cards phrased as questions on the front
• Unexpanded abbreviations
• Vague, essay-unfriendly backs
• Obvious common-sense cards that don't need studying
• Typos and grammar errors

Requires

• Anki 2.1.50 or newer, Qt6 build (the default and recommended version)
• An Anthropic API key from console.anthropic.com. You pay Anthropic directly for API usage — the add-on is free.

Setup

1. Install the add-on and restart Anki.
2. Open Tools → Claude Card Auditor → Settings.
3. Paste your API key. Adjust hotkey, model, or threshold if you want.
4. Start reviewing. Press Ctrl+Shift+A on any card.

Privacy

• Only the current card's Front and Back fields (stripped of HTML) are sent to Anthropic. No media, styling, tags, or deck metadata leaves your computer.
• Your API key lives in your local Anki config. Transmitted only as the x-api-key header to Anthropic.
• Audit history and feedback are stored locally in your Anki profile folder under claude_card_auditor/. Never uploaded anywhere.
• No telemetry. No ads. No bundled credentials.

Cost

Each audit is roughly 300–600 input tokens and up to 500 output tokens. At Claude Sonnet 4 prices, expect a fraction of a cent per card. Pick a cheaper model in Settings if you're auditing at volume.

Known limitations

• Currently reads fields 0 (Front) and 1 (Back) only. Cloze cards and unusual note types may not audit as expected.
• Applying a fix replaces HTML formatting (bold, italic) with plain text plus <br> line breaks.
• Scheduling reset uses the V3 scheduler (Anki's current default).

Support

Bug reports, feature requests, and card-design feedback welcome via the AnkiWeb page.