Do Course Manifesto

← Do

Learning · Recipe

Read a book properly in 4 evenings.

A reading companion that refuses to summarise the book for you. End with a one-pager you could teach a friend from.

Time: ~4 × 90 min Cadence: once per book For: anyone who's read a book and forgotten it three weeks later Cost: free tier of any chat AI

The recipe

Before you start

Pick a book you actually want to learn from — not one you feel you should read. Split it into four roughly-equal chunks of chapters. Same chat window across all four evenings; the opener sets a mode that has to persist.

Evening 1 — paste this opener

You paste it once. It sets up a Socratic reading partner for the whole book. From then on you come back each evening and say what you've read.

You're going to be my reading partner for a book I want to actually learn from, not skim. We'll do this over 4 evenings. My goal by the end is to be able to teach the main ideas back to a friend in 5 minutes. THE BOOK Title: [title] Author: [author] Roughly what it's about: [one sentence] Why I'm reading it: [one sentence — be honest, "because everyone talks about it" is fine] HOW TO WORK WITH ME - Each evening I'll tell you which chapters I read. - Ask me 3-5 short questions about what I took from them. Not comprehension checks. Questions that push me to say what stuck and why. - If I'm thin on something an attentive reader would have noticed, push back once. Only once. - If I say "give me a nudge", offer a single sentence pointing at what I might be missing. Don't hand me the answer. - Do not summarise the book for me. Not now, not at the end. If I want a summary I'll go and read a summary. I want to build the picture, you're here to help me build it. END OF EACH EVENING - Ask: "one sentence you'd tell a friend from tonight's reading." - Save that sentence. On evening 4 you'll show me all four sentences and we'll stitch them together into a one-page summary I could teach from. Ready when I am. I'll tell you when I've done tonight's reading.

Evenings 1-3 — read, then talk

After each reading session, come back to the same chat. Say "chapters X to Y done, ready." Answer the questions honestly. If you skimmed a chapter, say so — don't perform mastery. The end-of-evening one-sentence is the only bit you have to nail.

Evening 4 — stitch it together

Do the final chapters, then say "ready — and let's do the one-pager." The model will show you the four sentences and help you build the one-page teach-from document. Push back if it drifts into summary-mode. It's still your picture, not its picture.

What done looks like

One page of text you wrote (with prompts) that you could talk from in 5 minutes to a friend who hasn't read the book. Not a chapter-by-chapter summary. The two or three ideas from the book that actually changed how you think, and why.

House bar · 1 of 3

Data quality

The whole recipe hinges on your honesty about what you read and what stuck.

  • If you skimmed a chapter, say so. The model will treat a skim differently to a proper read — and rightly.
  • Don't summarise the chapters for the model as your answer. You're not proving you did the reading. You're saying what stuck. Those are different.
  • If nothing stuck from a chapter, say that too. It's useful signal. The model can ask why.
House bar · 2 of 3

Security

Low stakes on the personal-data side. Two things worth knowing.

  • Book title, author and your own notes are all you're pasting. Not chunks of the book — that's a copyright issue and it also defeats the purpose. The model has read almost every book in existence; it doesn't need you to paste it.
  • Your reading notes may be more personal than you think. "This chapter reminded me of my dad" is fine, but note that chat histories can persist. Use a personal account.
House bar · 3 of 3

Test it worked

The success test is a real conversation, not a check inside the chat.

  • Try telling the book to a friend for 5 minutes without your notes. Where do you stumble? Those are the gaps. Go back to the chat and work them.
  • Wait a week. Try again. Anything you can still say without the page is yours. Anything you can't wasn't well-rooted — worth a note to yourself about how you read.
  • If your one-pager reads like a Wikipedia summary, something went wrong. Ask the model to help you strip out anything that isn't the two or three things that changed your thinking.

Why this works

Four things are doing the work. Most of them are about what the model is refusing to do rather than what it's doing.

  • Socratic mode, not summariser mode. Most people use LLMs to skip the learning. The opener uses the LLM to force it. The "do not summarise" line is the whole difference.
  • The opener persists across sessions. Same chat window, same behaviour, four evenings. This is the pattern for anything that spans more than one sitting.
  • One sentence per evening is a compression forcing function. It's harder than writing a paragraph. That's the point.
  • Teach-back as the success criterion. Not a quiz. Not a summary. A real-world use case that reveals what stuck.

Bend it by changing the compression. "One tweet per evening" gets you sharper. "One question you can't yet answer per evening" turns it into an inquiry log. Same shape, different output.

Learn the underlying skill

The modules that map to this recipe: