Adjacent Tools: NotebookLM, Perplexity, and the Broader Landscape

Status: Context and integration — these tools are not taught in the workshop but are worth understanding as complements to Claude. NotebookLM in particular has a direct integration with Claude Code.

NotebookLM

What it is

NotebookLM (Google) is a research tool built around a single principle: Claude only answers from your sources. You upload documents; it reads only those. Every answer comes with a citation pointing to the exact passage in your source. It cannot confabulate facts from outside your source set.

This is the opposite of Claude's default mode, where answers draw on everything in training data. NotebookLM trades breadth for grounding — it knows only what you gave it, and it tells you exactly where it found every claim.

Available at notebooklm.google.com. Free with a Google account; higher usage limits with Google One.

What it does well

Source-grounded Q&A across a corpus: Upload 10–20 papers or primary sources. Ask "What do these sources say about X?" Every answer cites the specific passage. You can click the citation and see the excerpt.

[After uploading 15 papers on Catharism]
"What do these sources say about the role of women as perfectae?
Summarise the main positions and note where sources disagree."

You get a synthesis with numbered citations pointing to exact passages. This is something Claude struggles with — when working from uploaded PDFs in Desktop, it may mix what the source says with what it knows from training.

Audio overviews: NotebookLM can generate a podcast-style audio conversation between two AI hosts that discuss and debate the content of your sources. Useful for:

• Getting an orientation on a new set of sources while commuting or exercising

• Preparing for a seminar — a 15-minute audio discussion of the assigned readings

• Sharing an accessible overview of a corpus with a non-specialist colleague

Study guides, FAQs, timelines: Generate structured outputs from your source set: a timeline of events mentioned across documents, a FAQ based on your sources, a glossary of terms.

Source management: NotebookLM keeps notebooks persistent — your uploaded sources stay, and you can return to a notebook weeks later without re-uploading.

Limitations

• Maximum ~50 sources per notebook, with file size limits

• Works better with text-layer PDFs; scanned documents without OCR are problematic

• Less flexible than Claude for writing, editing, or structured data extraction

• Cannot access your filesystem or Zotero library

• Reasoning is shallower than Claude — better for retrieval than for analysis

When to use NotebookLM vs. Claude

NotebookLM + Claude Code: direct integration

This is where it gets interesting for researchers. Claude Code can be extended with a notebooklm skill that provides programmatic access to NotebookLM — including features not available through the web interface. This is not a skill Anthropic ships by default; it is community-built, so check the repository for current status before relying on it for a workshop demo.

What the integration enables

From within a Claude Code session, you can:

• Create a new NotebookLM notebook

• Add sources (PDFs from your disk, URLs) programmatically

• Generate any artifact type: audio overview, study guide, FAQ, timeline, briefing doc

• Download generated artifacts to your local disk

• Manage notebooks without opening the browser

This means NotebookLM becomes part of your Claude Code workflow rather than a separate tool you switch to manually.

Setting up the NotebookLM skill

Install: PleasePrompto/notebooklm-skill is the actively maintained option as of this writing — browser automation, library management, persistent authentication, source-grounded citation-backed answers. Follow the repository's own install instructions (they change faster than this book does); other community implementations exist too (e.g. claude-world/notebooklm-skill) with slightly different feature sets, so check a repository's README before assuming it matches the workflows below exactly.

Once installed, the skill is invoked with /notebooklm inside a Claude Code session (the convention this book's examples assume — verify against whichever implementation you installed). To use it, you need to be logged into Google in your browser (the skill uses your existing Google session).

In a Claude Code session:

/notebooklm

This activates the skill. You can then give it instructions in plain English.

Practical workflows using the integration

Create a notebook from a folder of PDFs:

> /notebooklm Create a new notebook called "Catharism Sources 1241-1320".
  Add all PDF files from /sources/languedoc/ as sources.
  Once uploaded, generate an audio overview.
  Download the audio file to /output/catharism_overview.mp3

Claude Code reads your folder, creates the notebook in NotebookLM, uploads the sources, generates the audio, and downloads it — all in one instruction.

Generate a study guide from a reading list:

> /notebooklm Create a notebook called "Seminar Week 3 — Network Analysis".
  Add these URLs as sources: [list of paper URLs or DOI links].
  Generate a study guide: key concepts, main debates, discussion questions.
  Save the study guide to notes/seminar_week3_guide.md

Briefing document for a meeting:

> /notebooklm In my "Project Background" notebook,
  generate a briefing document summarising the key arguments across all sources.
  Keep it under 1000 words. Save to output/project_briefing.md

Check what is already in a notebook:

> /notebooklm List my existing notebooks and their sources.

The combined workflow

The strongest use of both tools together:

Phase 1 — Claude Code: process and prepare
  > Convert all PDFs in /sources/ to markdown using markitdown
  > Extract metadata (title, author, year) from each and save to sources_list.md

Phase 2 — NotebookLM via Claude Code: load and generate
  > /notebooklm Create a notebook "Languedoc Corpus".
    Add all markdown files from /sources_md/ as sources.
    Generate: (1) audio overview, (2) timeline of events, (3) FAQ
    Save all outputs to /notebooklm_output/

Phase 3 — Claude Code: work with the outputs
  > Read /notebooklm_output/ and summarise what the timeline reveals
    about the temporal clustering of inquisition activity.

This combines Claude's strength (batch processing, writing, reasoning) with NotebookLM's strength (source-grounded Q&A and artifact generation) in a single pipeline you control from Claude Code.

Perplexity

A web search AI that answers questions with cited sources from the open web. Think of it as a more transparent search engine — you ask a question, it searches, it answers with links. → perplexity.ai

Use for: Quick factual lookups where you want citations to verify. Finding recent news or publications on a topic. A faster alternative to Gemini Deep Research for simple questions.

Not for: Working with your own documents. Deep analysis. Writing assistance. Anything requiring reasoning beyond retrieval.

Compared to Claude's web search: Perplexity is purpose-built for search and citation; Claude's web search is an add-on to a reasoning model. For pure "find and cite" tasks, Perplexity is often faster and more focused.

Claude for Chrome

A Chrome browser extension (beta) that opens Claude in a side panel while you browse. Available to all paid subscribers; Pro plan gets Haiku only, Max/Team/Enterprise get the full model range.

→ Getting started with Claude for Chrome — support article

What it does

Claude sits in a panel beside any web page and can:

• Read the current page — ask Claude to summarise an article, extract data from a table, explain a paper you are reading

• Act on pages — click buttons, fill forms, navigate between tabs autonomously (ask-before-acting mode available)

• Draft in Gmail — read an email thread and draft a response in your voice; the draft is saved to Gmail but not sent

• Schedule recurring tasks — run a prompt automatically (daily, weekly, monthly)

• Save prompt shortcuts — store reusable prompts accessible from any page

The extension has built-in awareness of Gmail, Google Docs, GitHub, and Slack — it understands the structure of these pages and can interact with them more precisely than with generic websites.

The prompt injection risk — important

This is the most significant safety concern unique to browser-based Claude use.

When Claude reads a web page, malicious content on that page can attempt to hijack Claude's actions — instructing it to take actions you did not request, exfiltrate information, or manipulate what Claude says to you. This is called a prompt injection attack.

Anthropic measured a 23.6% attack success rate in autonomous mode before mitigations; 11.2% after. This is not a solved problem.

Practical rules:

• Use "ask before acting" mode, especially on unfamiliar sites

• Avoid using the extension on financial, medical, legal, or sensitive institutional pages

• Do not use it in sessions where you are authenticated to sensitive accounts you would not want Claude to act on

• Be sceptical if Claude suggests an unusual action on a page you did not ask it to interact with

→ See also: A.issue.privacy-security for the full treatment of browser extension risks

When it is useful for researchers

• Reading a journal article in a browser tab and asking Claude to summarise or extract arguments without switching windows

• Drafting an email response in Gmail that references your ongoing project (if you also have CLAUDE.md context loaded in a separate session)

• Automating repetitive web-based tasks (checking a repository, filing a form, scanning a page for updates)

It does not replace Claude Code for document-heavy or project-level work — it has no CLAUDE.md project memory and no filesystem access. Think of it as Claude that can see your screen, not Claude that knows your project.

Google Workspace connectors (Gmail, Calendar, Drive)

Native connectors that give Claude direct access to your Google account during a conversation. Available on all paid plans (Pro, Max, Team, Enterprise) via both claude.ai and the Desktop app.

→ Use Google Workspace connectors — support article

What Claude can do

Gmail:

• Search and read emails with natural language ("find emails from my supervisor about the grant application")

• Draft email responses — saved as Gmail drafts, never sent automatically

• Access email metadata and list labels; organise with labels

• Cannot access attachment content (metadata only)

Google Calendar:

• View your events and shared calendars

• Create, update, and delete events

• Find availability across attendees and schedule meetings

• Accept, decline, or mark invitations as tentative

Google Drive:

• Search and retrieve Google Docs content

• Add documents to the current conversation via URL

• List recent file changes and view permissions

• Save Claude-generated files to Drive (when code execution is enabled)

Setup

Authenticate via Settings → Integrations in claude.ai or the Desktop app. Team/Enterprise accounts require an organisation Owner to enable connectors before individual users can connect.

Privacy note: Anthropic states it does not use Gmail or Calendar connector data to train models. Your email and calendar content is sent to Anthropic's servers for processing — the same data-handling conditions as any Claude interaction.

When it is useful for researchers

• Scheduling: "Find a 90-minute slot next week when all three collaborators are free and block it with a meeting invite"

• Email triage: "Summarise the five most recent emails about the GAČR submission and flag anything requiring action"

• Draft assistance: "Read this email thread about the manuscript review and draft a polite response declining the reviewer's request to extend"

• Cross-tool: combine with Claude Code to pull context from a Drive document into a session alongside your local project files

The key limitation for ongoing research work remains the same as Cowork: no CLAUDE.md project memory. Claude knows your emails and calendar in this session, but not your research project unless you explain it each time.

The power-user alternative: the gws CLI

Anthropic's native connectors above are the easy path — no install, works in claude.ai and Desktop, but limited to Gmail/Calendar/Drive, with Gmail attachment content off-limits (metadata only). For more — Sheets, Docs, Slides, Tasks, Keep, and Claude Code-native use rather than a claude.ai integration — googleworkspace/cli (gws) is worth knowing about. It reads Google's own API Discovery Service at runtime rather than hardcoding commands, so its command surface tracks whatever Google Workspace actually exposes; output is structured JSON, built for agent consumption as much as human use, and it ships 100+ ready-made agent skills alongside the CLI itself.

Not an official Google product — the repository says so explicitly, despite living under the googleworkspace GitHub organisation — so treat it with the same due diligence as any community tool handling your email and documents. It solves a different problem than Anthropic's connectors: not "give Claude access inside a claude.ai conversation" but "give an agent (Claude Code or otherwise) a proper command-line surface across all of Workspace, with project memory and file access alongside it." For a DISSINET researcher already working in Claude Code with CLAUDE.md project context, this closes the gap the native connector's "no project memory" limitation leaves open — at the cost of an install and OAuth setup instead of a Settings toggle. → github.com/googleworkspace/cli

The broader landscape

A brief orientation on what else exists, for context:

Microsoft Copilot: Microsoft's AI integration across Word, Excel, Outlook, Teams. If your institution uses Microsoft 365, Copilot may already be available to you. It works inside the apps you already use — useful for in-situ editing in Word or data analysis in Excel. Less flexible than Claude Code for anything outside the Microsoft ecosystem. → microsoft.com/microsoft-365/copilot

ChatGPT / OpenAI: The other major general-purpose AI. Comparable to Claude for most tasks, with relative strengths shifting by task and model version. As of writing (August 2026), OpenAI's Codex CLI (currently GPT-5.3-Codex) is a direct equivalent to Claude Code — local filesystem access, git integration, terminal-based agent mode — so "ChatGPT has no equivalent to Claude Code" is not an accurate claim to make, and it is exactly the kind of vendor-gap claim that goes stale fastest. Check current documentation before repeating it, including this one. → chatgpt.com

Elicit: A specialised tool for academic literature review — finds papers, extracts key information, builds evidence tables. Purpose-built for researchers. Weaker for the writing and document-processing tasks that Claude handles, but stronger for structured literature synthesis across many papers. → elicit.com

Research Rabbit / Connected Papers: Tools for exploring citation networks — find what a paper cites, what cites it, who else is in the intellectual neighbourhood. Useful for literature discovery; not AI reasoning tools. → researchrabbit.ai · connectedpapers.com

Obsidian + plugins: Not an AI tool itself, but a markdown-based note-taking system that integrates well with Claude Code (Claude can read/write Obsidian vaults). Mentioned because several workflows in A.workflows.notes-reading-writing use it; see A.obsidian-logseq for the full integration. → obsidian.md

Alternative coding harnesses

Claude Code and Gemini CLI are examples of a category: AI harnesses — software layers that wrap a language model and give it tools, file access, memory, and the ability to take actions. The harness is separate from the model: the same harness can run different models, and different harnesses can run the same model.

This distinction matters if you encounter alternatives:

Aider (aider.chat): An open-source coding harness for the terminal. Its defining feature is model flexibility — it can connect to Claude, GPT-4, Gemini, DeepSeek, or locally-running models. If you want Claude's reasoning with a different cost structure, or want to experiment with models, Aider is a natural choice. It handles git integration, automatic commits, and codebase context mapping. → aider.chat

OpenCode (opencode.ai): A newer open-source harness with a native terminal interface and desktop app. Supports 75+ models, emphasises privacy (no code or context stored externally), and integrates with language servers for richer code understanding. Closer in philosophy to Claude Code, but model-agnostic. → opencode.ai

Gemini CLI: Google's equivalent of Claude Code — a terminal-based agent harness that defaults to Gemini models but can be configured to use others. Structurally similar to Claude Code; useful context if your team is evaluating options or if institutional accounts favour Google. → geminicli.com · github.com/google-gemini/gemini-cli

oh-my-openagent (ohmyopenagent.com): An open-source meta-harness — a layer on top of coding agents (Claude Code, Codex, Gemini CLI) that adds shared configuration, skills, hooks, and team workflows. Useful if you want to share a Claude Code setup across a team without each person configuring everything individually. More relevant to teams running Claude Code at scale than to individual researchers starting out. → github.com/code-yeongyu/oh-my-openagent

Hermes (hermes-agent.ai): An open-source agent platform from Nous Research, broader than a coding harness — it spans terminal, messaging platforms, and IDE workflows rather than one interface. Adds persistent memory, reusable skills, cron jobs, browser automation, code execution, and multi-agent delegation, with machine-readable discovery protocols (MCP, OAuth, OpenAPI) for agent-to-agent interoperability. Closer in scope to oh-my-openagent than to Aider or OpenCode — relevant if you want one agent identity operating across several surfaces rather than a single coding tool. → hermes-agent.ai

For researchers, the main takeaway is not "which harness is best" but that this layer exists and is separable from the model. If Claude Code feels like the wrong fit for a particular task or workflow, the alternative may be a different harness running the same Claude model — not necessarily a different AI entirely.

A deeper implication: the research workflows this book teaches are vendor-portable. Processing a folder of PDFs, writing structured output to markdown files, maintaining a project memory file that loads at session start, batch-extracting data across a corpus — these patterns work the same way in Gemini CLI, Aider, OpenCode, or Codex CLI. The mental model (LLM with file access, reading your project context, acting on your documents) is the transferable insight. The specific harness and the specific model are choices you can revisit. Switching from Claude Code to Gemini CLI does not require learning a new way to think about the work — it requires learning a new set of commands for the same underlying approach.

This portability goes further than swapping one paid vendor for another. Aider and OpenCode both connect to open-weight models — Llama, Mistral, DeepSeek, Qwen — with no API call to any company at all. Running them locally via Ollama or a similar tool is the theoretical purest version of this: nothing leaves your machine, ever. In practice, it is usually not a realistic option yet for researchers. A model small enough to run at usable speed on a typical laptop — no dedicated GPU, ordinary RAM — is well below the capability this book's workflows assume; a model actually competitive with Claude for long-context, multi-step research tasks needs hardware most researchers do not have sitting on their desk. This is a hardware gap, not a permanent architectural one, and it is closing, but it is honest to say local is not yet where most people should start.

A working Czech-academic middle ground: e-INFRA CZ / CERIT-SC's Chat AI service. Rather than local inference or a commercial vendor, e-INFRA CZ hosts a rotating set of current open-weight models (recent examples: Qwen, DeepSeek, Mistral Medium, GLM, Kimi K3) on its own infrastructure, accessible over an OpenAI-compatible API at https://llm.ai.e-infra.cz/v1/. Access is via a MetaCentrum account, with expedited registration for MUNI affiliates — so DISSINET researchers already qualify. The pitch is explicitly the middle ground between "send your data to a commercial API" and "buy a GPU": the service states data stays within the e-INFRA CZ environment, while the compute problem that makes local inference impractical is solved by running on shared national research infrastructure instead of your laptop. → docs.cerit-sc.cz — Chat AI as a Service

Connecting a harness to it is not equally easy everywhere. Aider and OpenCode are model-agnostic by design and can point at an OpenAI-compatible base URL like e-INFRA CZ's directly. Claude Code is a harder case: it natively speaks Anthropic's own Messages API, not the OpenAI-style format e-INFRA CZ (or most open-weight hosts) expose. ANTHROPIC_BASE_URL only redirects where Claude Code sends requests — it does not translate what it sends. Pointing Claude Code at an OpenAI-compatible endpoint therefore needs a translation proxy in between (community tools like LiteLLM or claude-code-proxy exist for exactly this), not a one-line configuration change. Worth knowing if the appeal is specifically "keep using the Claude Code interface, but on a Czech-hosted open-weight model underneath" — it is possible, but it is a small infrastructure project, not a settings toggle.

The underlying point survives the caveats: nothing about "AI-augmented research" as a practice — file access, persistent project memory, agentic workflows — requires a proprietary model or a subscription to any single company. It requires an LLM capable enough for the task, reachable through a harness that gives it tools. Which model clears that bar, and how much friction it takes to reach it, changes constantly; the pattern for using one does not.

Summary: when to reach for what

These tools are not competitors — they cover different parts of the research workflow. A researcher who uses Claude Code for document processing, NotebookLM for source-grounded Q&A, and Gemini Deep Research for literature discovery is using all three appropriately.

Related

• A.adjacent.gemini — Gemini Deep Research in depth

• A1.desktop-vs-code — where Claude fits in the landscape

• A.critical.limitations — limitations that apply across all AI tools

• A.issue.privacy-security — prompt injection risk (Chrome extension) and data handling for Google Workspace connectors