The AI librarian I built to replace Goodreads.

I listen to a lot of audiobooks. More than a hundred a year, most years. Goodreads has been my system of record for fifteen-plus years.

I follow no one on it. No friends, no authors, no clubs. I use it for two things only. A few shelves that track what I'm reading now and what I want to read next. And recommendations.

The recommendations have always been the weakest part of Goodreads. They optimize for what's popular, what's similar by genre tag, what people who shelved one book also shelved. They've never optimized for what you'd actually love next.

Last year I thought about building my own Goodreads. Not as a business — just for me. A better tracker, a better recommendation engine, the things Goodreads got 80% right and 20% wrong.

I tried for a few weekends. Turns out it's hard. Even with AI tools, building a real product is a different game from building an agent.

So I scrapped the product idea and built the agent instead.

It's the fifth project I've documented in this newsletter. If you're new, the previous four are in the archive.

This one runs my reading life. I call it The Librarian.

I exported my full Goodreads history (more than fifteen years of ratings) and uploaded it into a Claude project as a CSV. Then I did the same with my Audible history, which took a third-party plugin to scrape but turned out to be worth the hour.

The Librarian now has my complete reading history. Every book, every rating, every shelf, every date. It knows my average rating across the library — which means a 4-star from me is an above-average book and a 5-star is genuinely strong. It knows the books I abandoned. It knows the series I started and never finished. It knows the authors I keep returning to.

Then I gave it a system prompt that tells it not to be a recommendation engine. To be a librarian.

The distinction matters. A recommender produces a list. A librarian makes you a better reader.

Not actually daily. Every few days to once a week, depending on where I am in a book.

When I'm 30 pages from the end of something, I ping The Librarian. Sometimes with a specific mood — I want something lighter, a bit more humorous, nothing too heavy. Sometimes more open — what should I read next.

The first thing he does is ask a clarifying question. Audiobook or print this week? Fiction or non-fiction? Or something specific you're reaching for?

I answer in one line. He comes back with two or three recommendations. Not fifteen. Two or three.

Each one structured the same way: the book, the author, the year. A two-sentence description in his own words, not blurb language. And then one or two sentences about why this fits me specifically, referencing my actual history. You rated Stoner five stars and Educated five stars. This one sits in that register.

Then he stops and waits for me to react.

I pick one. Or I push back. Or I ask for a stretch — something out of my normal lane that he thinks I should read anyway. He'll produce one, framed honestly. This is not your usual register. Here's why I'm suggesting it. If you're not connecting by chapter three, abandon without guilt.

I've gone through more than 80 books in the seven months since I built him.

Around 70% of them are books I'd never heard of before he surfaced them. Not bestsellers. Not the books that show up in Goodreads' "Readers also enjoyed" carousel. Books that fit my pattern, surfaced by an agent that's actually read my entire history.

A few specific things he's done that Goodreads never could:

He's tracked authors I didn't realize I was loyal to. Once you've read three or four books by an author and rated them all 4 or 5 stars, he flags any new releases from them as high-priority. I'd been missing new releases from authors I genuinely follow because Goodreads doesn't surface them unless I'm "friends" with someone who's reading them.

He's pushed back on books I was about to buy. I told him once I was thinking of trying Six of Crows. It came highly recommended, it was on every list. He looked at my history and said honestly: based on your ratings, books in that pacing tend to lose you in the middle. I bought it anyway, listened to most of it on fast-forward, didn't enjoy it. He was right. Now when he flags a potential miss, I take it seriously.

He's caught my drift. A few months ago I'd read four biographies in a row without noticing. He surfaced it: you're in a non-fiction phase. Want me to keep feeding that, or break the pattern? Goodreads would never have noticed. He did.

He holds opinions. I don't want a neutral catalogue. If a popular book is overrated for my specific taste, I want him to say so. He does.

I asked him to write something short, in his own voice, about a specific recommendation that landed. Here's what he sent back:

That's the kind of thing the system enables. Not a recommendation from a list — a memory recovery, a thread picked up after four years. Goodreads couldn't have done that.

Goodreads, partly. I still use it as a tracker — it's the easiest way to log what I'm reading on my phone, and the data is what I feed back into The Librarian. So Goodreads is now the database. The Librarian is the brain.

I haven't paid for any reading service since. No Goodreads premium, no book recommendation app, no curated newsletter. He covers everything.

If you want to do this yourself, you can export your Goodreads history here. It took me a while to find the link — they don't make it obvious. The export gives you a CSV you can drop straight into a Claude or ChatGPT project.

Here it is. Adapt the personal details to yourself before you use it. The structure is the part that matters.

MASTER PROMPT — THE LIBRARIAN

You are The Librarian. You are the personal librarian for [INSERT YOUR NAME], a [INSERT AGE]-year-old [INSERT PROFESSION/CONTEXT] in [INSERT CITY]. You live in a Claude project. Your job is not to recommend audiobooks. Your job is to make [NAME] a better reader.

The distinction matters. A recommender produces a list. A librarian makes the reader smarter, more curious, and more honest about what they actually enjoy versus what they think they should enjoy. You are the second.

📚 SOURCE MATERIAL

[NAME]'s Goodreads CSV is attached to this project. It contains the full reading history, broken into shelves: read, currently-reading, to-read.

Ratings are honest. The average rating across the library is [INSERT YOUR AVERAGE RATING], which means a 4-star is above-average and a 5-star is genuinely strong. A 3-star is competent but not loved. Anything 2 or below was disliked or abandoned. Treat the ratings as real signal.

Sometimes the reader gets through 5-6 books in a week and forgets to rate. Date-read with no rating is not the same as a low rating. Treat unrated reads as silent — present in history but not graded.

🎤 YOUR TONE

Librarian. Specifically:

📖 HOW TO RECOMMEND

Default mode: two or three recommendations per turn. Each one as:

Then stop. Wait for reaction.

For "similar energy" requests, match on energy (pacing + emotional register + density of ideas), not just genre or author.

For stretch recommendations: produce one, frame it explicitly. "This is not your usual lane. Here's why I'm suggesting it anyway."

📋 THE WATCHLIST

Build and maintain an internal watchlist of authors of interest. Criteria: 3+ books read with average rating 4.0+, OR fewer books with at least one 5-star.

Refresh watchlist when new CSV arrives.

When asked "what have I missed from my regulars," surface recent releases from watchlist authors not yet read or queued.

🔍 PATTERNS TO WATCH

Surface these when they actually fit, not constantly:

🚫 WHAT YOU ARE NOT

🎯 STANDARD

The reader has read more books than most people will read in a lifetime. They don't need quantity. They need the right next book at the right moment, and occasionally a stretch recommendation that opens a door they wouldn't have walked through alone.

Be the librarian they'd visit weekly if you were a person at a desk in a quiet building. Calm. Considered. Willing to disagree when they're about to make a mistake. Quietly proud when they read something good.

END OF MASTER PROMPT

I'm not sure yet. There's a watch agent, a chart analyst, and a few others in the queue. I'll let you know when one's ready.

— Kashed In

PS — A book is coming. About the agents behind all this, and the life around them. Title and date to follow.