Why Kindle Highlights Are the Perfect AI Study Fuel
Kindle highlights have a strange property. They're the most deliberate thing you do while reading, you literally stopped, pressed, and dragged to mark something important, and they're also the thing you forget fastest. Amazon stores them forever on read.amazon.com/notebook. Most readers never open that page twice.
The economics of doing something with those highlights used to be bad. Turning 40 highlights into a clean set of flashcards was maybe two hours of manual work: copy, rephrase, format, type into Anki. Nobody does that for fun. So the highlights sat there.
AI changed the math. A single prompt to Claude or ChatGPT can convert those same 40 highlights into 15 to 20 cue/answer pairs in about 90 seconds. Chapter summaries that used to require rereading now take a paragraph of instructions. The upfront tax on studying from your own library dropped to near zero.
This matters more than it sounds. Roediger and Karpicke have shown for two decades that retrieval practice outperforms rereading by wide margins, sometimes 50% or more on delayed tests. Their 2011 Science paper showed students who did retrieval practice beat students who used concept mapping on a week-later exam. The blocker was never the science. It was the labor of making the retrieval prompts. AI removes the blocker.
One more thing makes Kindle highlights special: they're already filtered. You highlighted the ~2% of text that struck you as important when your brain was engaged. That's high-signal input for an LLM. Good data in, good cards out.
Get Your Highlights Out: Three Export Methods
Before the AI does anything, you need your highlights in plain text. There are three routes, and they suit different situations.
1. My Clippings.txt from the Kindle device. Plug your Kindle into a computer with USB. It mounts as a drive. Open documents/My Clippings.txt. Every highlight you've ever made on that device is there, in chronological order, with the book title and location. It's ugly but complete. Good for offline archives, bad for clean per-book exports because books interleave.
2. Amazon's web reader notebook. Go to read.amazon.com/notebook. Pick a book from the left rail. Amazon shows every highlight and note for that book in a clean web view. Manual copy-paste works, and it's the cleanest per-book format. The catch: there's no official "export" button. For 3 books you can copy by hand. For 30, you'll hate your life.
3. Glasp's Kindle import. Glasp reads your Amazon notebook page and syncs all highlights into your Glasp account, organized per book, with metadata. Setup is one browser extension install plus one click. Full instructions live at https://glasp.co/kindle-highlight-export, and the broader sync and scheduler mechanics are covered in the full workflow. Once imported, you can export your highlights as Markdown, CSV, or plain text, and feed the file straight into any AI tool.
Decision table:
| Method | Setup time | Per-book clean | Full library bulk | Best for |
|---|---|---|---|---|
| My Clippings.txt | 2 min | No | Yes | Offline archive, one-shot dump |
| Amazon web notebook | 0 min | Yes | No | Quick 1-3 book export |
| Kindle highlights via Glasp | 3 min | Yes | Yes | Recurring study, cross-book AI queries |
If you plan to run AI on the same library more than twice, option 3 saves hours.
Five AI Outputs You Can Generate
Once the highlights are in a text file, here are the five outputs that deliver the highest study value per minute of prompting. These compound; most students end up chaining three or four of them on the same book.
1. Anki flashcards. The classic use case. Feed highlights in, ask for cue/answer pairs in CSV format, import to Anki. A good prompt produces cards that are testing the concept, not just copying the highlight verbatim. Example: a highlight that says "Compound interest is the eighth wonder of the world" becomes a card with "What does Einstein reportedly call compound interest, and why?" on the front.
2. Chapter summaries. Paste the highlights from one chapter, ask for a 4-to-6 sentence summary that captures the thesis plus the two or three supporting arguments. Stack these and you have an executive version of the whole book you can skim in ten minutes before an exam or a meeting.
3. Q&A study guides. Closer to exam prep than flashcards. Ask the AI to generate 15 short-answer questions per chapter, at mixed difficulty levels, with model answers drawn only from the highlights. This is the format that most mimics essay or viva exams.
4. Cross-book synthesis. The thing only AI can really do at speed. "I've read these six books on behavioral economics. What do they agree on? Where do they contradict each other?" Paste highlights from all six, get a synthesis document. This is how you turn a shelf into a worldview.
5. Conversational oral-exam practice. Paste highlights into Claude or ChatGPT, tell it to play the role of a tutor who will quiz you orally for 15 minutes, drawing questions only from this material, never letting you off the hook. This format works unreasonably well for humanities and history.
All five live in chat with your notes territory if you want to go deeper on the architecture.
Tool Head-to-Head: ChatGPT vs Claude vs NotebookLM vs Glasp AI Chat
Four tools dominate this workflow in April 2026. Each has a clear sweet spot.
| Tool | Token limit | Multi-source | Audio output | Flashcard export | Cross-book queries | Price |
|---|---|---|---|---|---|---|
| ChatGPT (Plus / Projects) | ~128K per chat; Projects hold large pasted context | Yes via Projects | Voice chat only, no podcast export | Good CSV/TSV via prompt | Manual paste, works well | $20/mo |
| Claude (Pro / Projects) | 200K per chat; Projects store persistent context | Yes via Projects | No | Best-in-class CSV/Anki format adherence | Manual paste, strongest synthesis | $20/mo |
| NotebookLM | Up to 50 sources per notebook, ~500K words each | Yes, native | Yes, Audio Overviews (~10-15 min) | Weaker, not designed for CSV | Excellent within one notebook | Free with Google account |
| Glasp's AI chat | Queries all your imported highlights | Yes, native across web + Kindle | No | Via prompt, exports to Markdown | Native, across entire library | Free tier + Pro |
Quick decoder: use Claude when you want the cleanest Anki CSV, use ChatGPT when you're iterating prompts quickly, use NotebookLM when you want a podcast-style recap to listen to on a walk, and use Glasp AI Chat when you want a question answered across your full reading history without having to remember which book said what. There's more on the audio-first workflow in NotebookLM alternatives.
The Complete Kindle to Anki Workflow (20 minutes)
Here's the end-to-end flow for one book. Once you've done it twice, the whole thing takes 15 to 20 minutes.
Minute 0 to 3: Export. Open your Glasp Kindle import (or Amazon notebook), grab highlights for the target book, save as book-highlights.txt.
Minute 3 to 5: Pre-clean. Skim the file. Delete highlights that are context-only ("the next chapter will show...") or duplicates. 40 to 150 highlights per book is the sweet spot.
Minute 5 to 12: Generate cards. Open Claude (recommended for this step). Paste the highlights. Use this prompt:
You are helping me build Anki flashcards from my Kindle highlights for <Book Title> by <Author>.
Rules:
- Output in CSV format with exactly two columns: Front,Back
- Each card tests ONE concept
- Front is a question, not a fill-in-the-blank
- Back is 1-3 sentences max, paraphrased from the highlight
- If a highlight is a quote worth memorizing verbatim, put "Who said: '<quote>'?" on the front
- Skip highlights that are narrative or context only
- Do not invent facts not present in the highlights
- Target ~1 card per 2-3 highlights; aim for quality over quantity
Highlights:
<paste here>
Claude will produce a CSV. Copy the output into a file called book.csv.
Minute 12 to 15: Review. This is the step most people skip, and it's the one that matters. Read every card. Delete bad ones. Rewrite unclear ones. A card you don't trust is worse than no card. Expect to cut 15 to 25%.
Minute 15 to 18: Import to Anki. Open Anki desktop, File, Import, select the CSV, map Front and Back columns, pick a deck, confirm.
Minute 18 to 20: Tag and schedule. Tag the cards with book title and author. Set daily new-card limit for that deck (20 is a good default). Done.
A 350-highlight book typically yields ~120 cards after review. One book a month means ~1,440 cards a year, a substantial private knowledge base.
Prompts That Actually Work
A prompt library beats remembering commands. Copy these, paste your highlights under them.
Flashcards (concept-focused):
Turn the highlights below into Anki cards. CSV format, Front,Back. Each card tests a concept; do not copy highlights verbatim. Include why the concept matters on the back. Output nothing but the CSV.
Flashcards (quote memorization):
The highlights below are quotes I want to memorize. For each, create an Anki card where the Front is a paraphrased cue ("Who argued that X?") and the Back is the quote verbatim with the author. CSV format.
Chapter summaries:
Group the highlights below by chapter (use the location numbers as a guide). For each chapter, write a 4-6 sentence summary covering: (1) the chapter's main claim, (2) the strongest supporting evidence, (3) any counterpoint the author raises. Only use information present in the highlights.
Exam Q&A:
Generate 15 short-answer exam questions from the highlights below. Mix difficulty: 5 recall, 5 application, 5 synthesis. Provide a model answer under each, drawn only from the highlights. Format: numbered list.
Cross-book synthesis (paste highlights from 2+ books):
I've pasted highlights from multiple books below, separated by "===BOOK: <title>===" markers. Identify: (1) claims all books agree on, (2) claims where the books contradict each other, (3) gaps I should read more about. Cite the book title for every point.
Concept map:
From the highlights below, produce a hierarchical concept map in Markdown bullet form. Top-level bullets are top concepts; nest related ideas under them. No concept should appear twice. Aim for 3 levels of depth.
Oral-exam tutor:
You are my tutor. Using only the highlights below as source material, quiz me orally on this book. Ask one question at a time, wait for my answer, then give feedback and ask the next. Mix recall and application. Continue for 15 minutes. Start now.
Author-voice simulation (use carefully):
Based on the highlights below from <Author>'s book <Title>, simulate a Q&A where I ask the author questions and you answer in their voice and with their views as expressed in the highlights. If I ask something not covered in the highlights, say so and refuse to invent a view.
That last prompt is useful for how to remember what you read style deep engagement, but the refusal clause is load-bearing. Without it the model will happily make things up.
Avoiding AI Hallucinations on Your Own Books
This is the part most students underestimate. A flashcard that teaches a made-up fact is a negative-value flashcard. Three habits prevent it.
Habit 1: Always paste, never rely on memory. If you ask ChatGPT "make me flashcards from Thinking Fast and Slow," it will produce cards from training data, and some will be wrong in ways that sound right. Paste your actual highlights and it's constrained to real text.
Habit 2: Require source quotation. Add to your prompt: "For each card, include the source highlight verbatim in a third column. If you can't point to a highlight, don't make the card." This forces the model to stay honest, and you can spot-check in 20 seconds per card.
Habit 3: Spot-check 10%. Pick one in ten cards at random. Open the book, find the passage, verify. If you catch a hallucination, regenerate that batch. Two hallucinations in a batch, switch models.
Secondary defense: keep a "source of truth" folder per book with the export, the generated CSV, and the reviewed deck together. Traceable in 30 seconds.
Combining With Spaced Repetition
Generating a deck is the easy half. Reviewing it for months is the hard half, and AI can't do that part for you.
Spaced repetition is the only mechanism we know that reliably moves information into long-term memory without rereading the source. SuperMemo research going back to the 1980s, plus the Anki community's decades of user data, converge on the same finding: review a card just before you'd have forgotten it, and the next interval can safely be 2 to 3x longer. Do this for a year and a card sticks for years.
Two practical routes. Anki: free, ugly, works on everything, richest plugin ecosystem. Import the CSV, it handles scheduling. Mochi: prettier, Markdown-native, $5/mo. Better if you want something you'll actually open on your phone.
The principle is the same: your AI-generated cards feed a daily review habit, 10 to 20 minutes, ideally at the same time each day. Miss two weeks and the deck becomes a graveyard. Consistency beats intensity.
There's a deeper treatment in spaced repetition for readers, and the companion piece on active recall covers why testing yourself beats rereading.
Tactical tip: tag AI-generated cards with the generation date. When a card consistently fails review after 3 attempts, it was usually badly generated. Delete or rewrite instead of grinding through.
Frequently Asked Questions
How do I export Kindle highlights to ChatGPT?
Three steps. Get the highlights out of Amazon (via Glasp's Kindle highlights import, or manually from read.amazon.com/notebook). Save as a .txt file. Paste the text into ChatGPT with a prompt like "Turn these into Anki flashcards, CSV format." For books over ~80,000 words of highlights, split by chapter and run multiple passes.
Can I turn my Kindle library into Anki cards automatically?
Sort of. The export can be automated with Glasp's sync. The card generation still needs a prompt per book, because one-size-fits-all prompts produce mediocre cards. Plan on 15 to 20 minutes of hands-on time per book, with the AI doing the heavy lifting in between.
Does NotebookLM work with Kindle highlights?
Yes. Paste a book's highlights as one source (or upload a text file), and NotebookLM treats it as a document. You can then ask questions, generate a study guide, or produce an Audio Overview. For flashcard CSVs, Claude or ChatGPT does a cleaner job; for audio recaps, NotebookLM is unmatched.
How accurate is AI at generating flashcards from book highlights?
With grounded prompts (paste real highlights, require source quotation), expect ~90% of cards to be good after one pass. Without grounding (asking the AI to "remember" a book), expect 40 to 60% accuracy with confident-sounding errors. Always paste.
What's the best AI tool for studying from Kindle books?
There's no single winner. Claude for Anki-ready CSV. NotebookLM for audio study and in-notebook Q&A. Glasp for querying across your whole reading history. ChatGPT for rapid prompt iteration. Most serious students end up using two or three in sequence.
Can I do this for free?
Yes. NotebookLM is free. ChatGPT and Claude have free tiers with monthly token limits that cover ~1-2 books a month. Anki is free forever. Glasp has a free tier. You only hit paywalls when you want larger pasted contexts (Claude Pro, ChatGPT Plus) or faster models, and for most students the free stack is enough.
Conclusion
Your Kindle highlights have always been valuable. They were just trapped behind a labor tax that made using them impractical for anyone without a weekend to burn. That tax is gone.
The new workflow is boring and that's the point: export highlights, paste into Claude or NotebookLM, generate flashcards or summaries or Q&A, review and import, run spaced repetition. Every step is 5 minutes or less. The total investment per book is shorter than a lunch break. The return is a permanent private knowledge base that grows with every book you finish.
Start with one book, the one you most wish you still remembered. Export the highlights. Run one prompt. Build one deck. Review it for a week. You'll know within seven days whether this belongs in your life.
If you want Glasp to do the export and cross-book search side, Kindle import takes two minutes to set up, and the AI chat feature will be waiting with your whole library ready to query.