Claudefor a reading recommendation, it gives you a polished, slightly generic summary, and you think, "Sure, sounds interesting," only to realize fifty pages in that the prose is dry or the tone is completely off. The problem is that LLMs are great at summarizing, but summaries are inherently reductive. They strip away the soul of the writing—the rhythm, the specific voice, and the actual texture of the ideas.
I recently started testing a Claude connector called Coffeetable, and it fundamentally changes the workflow for anyone using LLMs as a personal research or discovery assistant. Instead of just getting a text-based description of a book, the tool allows Claude to pull actual pages from the text directly into the chat interface.
How the workflow actually looks #
When you have this connector installed, you aren't just chatting with a model that has "knowledge" of a book; you are interacting with a model that has "access" to the book.
- The Discovery Phase: You start with a high-level prompt. I’ve found that the most effective way to trigger deep, personalized results is to lean into the long-term memory aspect of Claude. Instead of asking "What should I read?", try a prompt like this:
Based on everything you know about my interests, my current projects, and the way I think, what are some books I must read next?
2.The Verification Phase: Once Claude suggests a title, you don't have to take its word for it. You can immediately follow up with a request to see the actual content.
The Deep Dive: You can ask for specific sections. For example: "Show me the first three pages of Chapter 2 so I can get a feel for the author's style."
Why this matters for prompt engineering #
This shifts the way we approach an AI workflow for learning. Traditionally, prompt engineering for book discovery was about refining the description of what you wanted. Now, it’s about refining the sampling of the source material.
By bringing the actual pages into the context window, you are performing a real-world validation of the LLM's recommendation. It turns Claude from a mere librarian into a preview reader. If the prose in those sampled pages doesn't click with you, you've saved yourself hours of reading a book that wasn't a good fit.
Practical tips for using connectors #
If you are building out your own LLM agent ecosystem, this is a perfect example of why "connectors" are the next big step. A model is only as good as its ability to ground its responses in real-world data.
Avoid generic queries: If you ask for "best sci-fi books," you'll get the same top 10 list from every AI.Use specific stylistic prompts: Ask Claude to "Show me a passage from this book that demonstrates its core philosophy" to see if the actual writing matches the hype.Combine with note-taking: Use the sampled pages to create immediate, high-quality study notes or summaries within the same thread.
It’s a massive jump in utility when you move from "tell me about this" to "show me this."
Next Stop making your stakeholders do mental math during eval reviews →
these AI tool field notes, with plenty of directly applicable cases.