@itsolelehmann: how to *actually* remember what you read (using Karpathy's LLM reading method): let's be honest...nobody remembers what…

how to actually remember what you read (using Karpathy’s LLM reading method):

let’s be honest…nobody remembers what they read anymore.

you finish a book, highlight a few good lines, feel smart for a week, then 95% of it disappears.

karpathy found a better way to read with AI.

a way that turns reading from passive information consumption into actual understanding that you remember.

it’s called reader3.

he built it so you can read EPUBs chapter by chapter with an LLM as your private tutor.

you read the book yourself first, then you leverage AI to turn what you read into working knowledge.

here’s how to run it:

1. download the EPUB into reader3

reader3 is basically a chapter slicer for books.

you drop in the EPUB, and it turns the book into clean chapter-by-chapter sections.

so after you read chapter 4, you can send only chapter 4 to your LLM instead of uploading the whole book or digging through the EPUB yourself.

for the actual reading, use whatever you like.

physical book, kindle, iPad, whatever.

i use a kindle.

2. pass 1: read manually

read the chapter yourself first in your normal reading setup.

no laptop, phone, or AI allowed.

just you and the book.

this gives you your own first-pass understanding before the model starts explaining things.

3. pass 2: have the LLM explain it

after you finish the chapter, open that same chapter in reader3 on your computer and send it to your LLM.

ask it to explain and summarize the chapter in plain english.

i like to prompt: “explain this chapter so a 15-year-old wouldn’t get lost”

this is where the LLM becomes a tutor sitting next to you.

it slows down the confusing parts, explains references, and makes the argument easier to see.

4. pass 3: Q&A

now pepper the chapter with any questions you have:

what did i miss? what are the 3 ideas worth remembering? what was the author actually arguing? where was the author weak or unclear? quiz me on the important parts connect this to the previous chapters

the questions force you to actually work and engage with the material instead of just collecting another summary.

5. save it to your wiki

after the Q&A, have the AI turn the chapter into notes:

chapter summary key ideas best quotes open questions themes links to previous chapters

save those notes into obsidian, a markdown folder, or an LLM wiki.

now the book compounds as you read.

chapter 8 updates ideas that started in chapter 2.

arguments get tracked across the whole book.

and by the end, you have a searchable private memory for everything you learned.

that’s the flow:

kindle for attention. LLM for private tutoring. wiki for long-term memory.

karpathy’s original tool: https://github.com/karpathy/reader3…

i recommend using LLMReader as the upgraded version of the same concept: https://github.com/yongkangc/llmreader…

it adds EPUB/PDF support, easier chapter copy, highlights, notes, and markdown export, so you can send everything into obsidian or your LLM wiki after each chapter.

karpathy’s original tweet:

Claude Code + Granola is a literal cheat code for making meetings actually productive.

the problem: most meeting transcripts die right after the call. all the decisions and action items just sit in a doc nobody opens again.

so i built a skill that does the after-meeting work nobody gets around to:

it turns your granola notes into an interactive dashboard that visually shows your team exactly what to do next…

so what you decided actually gets done.

here’s how it works

you connect granola to claude code or codex via MCP and after any call you just say:

“make a dashboard from my latest granola meeting”

the skill reads the transcript, figures out what kind of meeting you just had, then builds the right dashboard for it.

if it was a brainstorm, you get an idea board: the best ideas, the themes, the experiments to run next.

if it was a planning session, you get a timeline: milestones, owners, dependencies, blockers.

if it was a sales call, you get a one-pager you can send to anyone who missed it: the customer’s pain, their objections, exact quotes, next steps.

and if you keep an LLM wiki / second brain, it exports the important context there so claude remembers the decision later.

the point is simple: after a meeting you should know exactly what happened, what matters, what to do next.

agents handle the doing now. meetings are for humans deciding on what’s worth doing.

so the call ends, the agents grab the dashboard and run with it.

full skill below: