How to Turn AI Conversations Into Anki Flashcards for Spaced Repetition

A three-hour Claude thread on Spanish subjunctive, a long Gemini answer about kidney physiology, an annotated ChatGPT walk-through of Rust lifetimes. These are the AI conversations worth remembering, and the ones you forget the fastest. Spaced repetition is the cure, but the chat apps will not get you there on their own. The workflow below turns any saved AI thread into a clean Anki deck without retyping a single answer.

Why AI Threads Are a Great Source for Flashcards

A long AI conversation already contains the question and the answer. The prompt is the front of the card, and the model's reply is the back, often with the example, the worked counterexample, and the small aside that made it click in the first place. Most study material has to be rewritten before it becomes a flashcard. AI threads do not, because the format is already a question and answer log by construction. That structural match is the whole reason this workflow exists at all.

The problem is that the chat window is the wrong shape to extract from. You cannot multi-select across messages, you cannot mark a paragraph for export, and the moment you close the tab the thread becomes a URL with no offline copy. The chat app was built to write the next reply, not to retrieve the last one. If you want to study from these threads, the first move is the same one covered in save AI conversations you want to re-read: get the markdown out of the app and onto disk. Once it is on disk it is yours, and the rest of the workflow follows.

Export the Thread as Markdown First

Anki does not import a chat URL, and it does not parse HTML cleanly either. It wants a flat text source, and markdown is the cleanest one you will get from ChatGPT, Claude, Gemini, or Perplexity. Use the model's built-in export where one exists, and the copy-paste-into-Prism-MD route where one does not. The full per-model export routes are in how to read long Claude conversations and the model-specific posts linked from it. Pick the path that matches the model you reach for most days.

The output you want is one markdown file per study topic, with the prompt as an H2 and the reply as the body. If your export bundles many topics into one file, split it before you start picking cards, because Anki imports cleaner from focused sources. A Spanish grammar deck and a kidney physiology deck have nothing to do with each other, and Anki will let you merge later if you change your mind. Naming the files with a topic prefix saves you a sort later. The half-minute spent splitting at export time pays for itself the first review session.

Pick Cards That Are Worth Reviewing

The temptation is to turn every paragraph into a card. Resist it. A good flashcard has one prompt and one short answer, and an AI reply is usually four paragraphs long. The cards that survive a year of review are the ones that test a single fact, a single rule, or a single distinction. Everything else is reference material that belongs in the source document, not in the deck.

Read the thread once with a highlighter pass in mind. The same approach works here as in annotating AI conversations, where the goal is to mark the parts that surprised you, not the parts that confirmed what you already knew. A surprise is a candidate for a card, and a confirmation is not. If you are studying a language, the candidate is the irregular verb the model corrected you on. If you are studying anatomy, it is the artery you keep confusing with another one. Stay strict about this filter and the deck stays useful.

The kinds of moments worth a card usually fall into a short, repeatable set. They are the lines you underlined on the first read because they overturned something you thought was settled. They are the lines that filled a gap your previous notes had left open without you noticing. They are rarely the summary paragraph at the top of the reply, because summaries flatten the surprise that made the answer worth keeping. Keep this list nearby on the first few passes, and the deck will stay weighted toward signal:

• The correction the model made when you got something wrong.
• The exception to a rule the textbook stated as universal.
• The mnemonic the model offered that you remembered an hour later.
• The single number, date, or formula the rest of the answer hung on.

Aim for ten to twenty cards from a one-hour thread. That sounds low until you remember that Anki will surface each card hundreds of times across a year of review. Quality beats volume by a wide margin once a deck gets past a few hundred cards. The discipline of picking fewer cards is what keeps the deck reviewable instead of abandoned.

Convert the Markdown Into Anki Cards

Anki imports from two sources that matter here: a tab-separated or comma-separated file, and a markdown file through the right add-on. The simplest path is to write a tiny CSV with two columns, front and back, and import it with File then Import inside the Anki desktop app. The official format is documented in the Anki manual on importing and has not changed in years. That stability is part of why a CSV pipeline is the safe default for anyone starting out.

If you prefer to keep the source as markdown, the Markdown Format add-on reads files where each card is separated by a horizontal rule, and the front and back are separated by a single blank line. That format is friendly to a small script, and a small script is the right answer once you have more than one thread to convert. A ten-line Python script that walks a folder of markdown exports and writes one CSV is usually all you need. Keep the script in the same folder as the decks so you can rerun it without thinking.

Whichever path you pick, keep the source markdown next to the deck. A flashcard with no context is a trivia question, and a year from now you will want to find the original answer the card came from. The folder layout that works is a topic per folder, a thread per markdown file, and a single CSV or apkg next to them. That is the same archive layout that makes the search workflow further down work without ceremony.

Keep the Deck in Sync With New Threads

The hard part of any study deck is not the first hundred cards. It is the second hundred, six months later, when the topic has come up in three new conversations and the deck has not moved. The fix is a weekly habit, not a tool. Once a week, open the threads you saved that week, pick five to ten new cards, and import them into the same deck. The habit beats any clever automation because the picking is the part that requires judgment.

The search workflow in searching saved AI conversations is the right partner here. When a review card stumps you, search the archive for the original thread, reread the surrounding paragraph, and either rewrite the card or add a sibling card that tests the gap. The deck improves because the source library improves, and the source library improves because you reread it. That feedback loop is the whole point of keeping the source in a searchable folder rather than a closed chat tab.

FAQ

Does this work for image-heavy threads?

Partly. Anki supports image fields, and the markdown export from most chat apps inlines images by URL. You will want to download those images locally before importing, because hotlinked images break the moment the chat app rotates its CDN. Storing the images next to the deck file keeps the cards self-contained.

Can I share the deck with a study group?

Yes. Anki exports a deck as an apkg file that anyone can import, and AnkiWeb lets you publish a shared deck if you want it discoverable. Strip personal prompts from the front of the cards before sharing, because the prompt often contains your name, your role, or the specific exam you are studying for. A quick find and replace pass usually catches the obvious ones.

What if the AI got the answer wrong?

Catch it now, not in review. Reread every reply before you turn it into a card, and cross-check the model against a textbook source for anything that will be tested on an exam. A wrong card reviewed two hundred times is worse than no card at all. The five minutes spent verifying a fact is the cheapest insurance in the whole workflow.

Should I use Cloze deletion or basic cards?

Both, and the source thread will tell you which. A definition reply turns into a cloze, because the structure is already a sentence with a key term to hide. A worked example with a single answer turns into a basic card, because the prompt and answer are already separated. Mixing them inside one deck is fine, and Anki handles the scheduling for both the same way.