Show HN: Turn Newsletters into Interactive GPTs

[raunaqvaisoha]

I’ve been hacking on a project called Bookshelf (https://www.bookshelf.diy/). It lets you take an archive — say, your Substack export, a bunch of PDFs, or even saved HTML files — and turn that into a retrieval-backed GPT that your readers can query.

The idea is: instead of scrolling archives, they just ask questions. Answers are pulled only from your original content, with citations.

It’s aimed at writers and researchers who want their work to be more discoverable — but without spinning up vector infra or fiddling with RAG pipelines.

For context: I’ve always gone back to Paul Graham’s essays for startup advice. But there’s no good way to search them semantically or contextually. So I tried indexing a few with Bookshelf.

Asked: “How does PG think about evaluating founders?” and got a clean answer sourced from Do Things That Don’t Scale and a couple other essays — citations included. It was surprisingly useful.

So far, one early test case is AnthropoceneGPT (https://sammatey.substack.com/p/introducing-anthropocenegpt) for Sam Matey’s newsletter. It’s seen ~100+ queries. Readers say it works like a smart librarian. He says it gives him ideas for what to write next.

Rough implementation: Input: HTML/PDF exports Chunks + embeds via OpenAI (or local) Stored in a vector DB Retrieval API is called by the custom GPT GPT is instructed to only use retrieved chunks and cite them Auth Option: for tracking on queries to give writers some telemetry

Here’s a demo GPT trained on Paul Graham’s archive: Paul Graham GPT (https://tinyurl.com/paul-graham-gpt)

Would love thoughts on: What would make this better for writers or readers? Any UX nits on the GPT side? Has anyone tried doing something similar in-house?

[fbohs888]

The "smart librarian" analogy for AnthropoceneGPT rings true. As someone who's tinkered with RAG locally, the promise of avoiding the "spinning up vector infra or fiddling with RAG pipelines" is incredibly attractive. Really impressed with the concept!

One UX thought on the GPT side: how prominent are the citations? And is it easy for readers to click through to the original content directly from the GPT's response? Making that flow seamless would be a huge win for verifying information and deeper engagement.

[sunny9911]

This is really cool! It let me upload my documents and create a custom GPT. Now, anyone I share the link with can ask questions and get answers based only on what I’ve uploaded.

It’s like having a private assistant that only knows what I’ve written. Setup took some to and fro between ChatGPT and Bookshelf. I also love how it gives citations from the document so I can double check. Till now, it has not hallucinated. Great job bookshelf team.

[korgy]

This is pretty clever. I can definitely see the appeal for writers with big archives that readers don’t have time to sift through. I’m wondering though — does it handle more conversational queries well, or is it better for straightforward factual lookups?

[sahilkat]

It actually works well for conversational queries too. As long as the topic has been covered in the newsletters, it can handle both casual and direct questions. The responses are designed to reflect the author's own style, but it always sticks to what’s in the newsletters—so to avoid hallucination.

[soman3]

The telemetry idea is great, being able to see what people are querying could even inspire future essays/newsletters.