A12: Good Practice Examples — Claude Desktop Used Well for Research

Compiled from web research, March 2026. Sources verified.

The key insight: Desktop done well requires deliberate setup

Most researchers using Claude Desktop use it at the most basic level — single conversations, no persistent context, no Projects. The gap between basic Desktop use and well-configured Desktop use is large, and almost entirely a matter of setup rather than skill.

The examples below show what Desktop looks like when used deliberately.

Thesis writing: Obsidian + Zotero + Claude Desktop

"My Thesis Writing Workflow (Obsidian, Zotero and Claude AI)"

The workflow documented: 1. Reading (40–45% of time): Literature read with atomic note-taking in Obsidian — small, self-contained notes linked to Zotero entries 2. Synthesis: Claude Desktop used to synthesize across notes — "given these notes, what are the main tensions in the literature?" 3. Writing: Claude assists drafting sections, with Zotero citations already embedded in the note context

The key principle: Claude is not asked to do the reading. The researcher does the reading and makes the notes. Claude synthesizes across notes that already represent the researcher's processed understanding. This keeps interpretive judgment with the human.

NotebookLM vs. Claude Projects: choosing the right tool

Source: https://www.atlasworkspace.ai/blog/notebooklm-vs-claude-projects

This comparison is directly useful for researchers choosing between tools:

Task Better tool Why
"What do my sources say about X?" NotebookLM Strict source grounding; inline citations; lower hallucination risk
"Synthesize these sources with broader context" Claude Projects Draws on training knowledge beyond uploads; deeper reasoning
"I need verifiable quotes with page references" NotebookLM Citations are its core feature
"Help me develop an argument across sources" Claude Projects Better at novel synthesis
"I have 50 sources to work through" NotebookLM Handles large source sets cleanly
"I need Claude to also write and edit" Claude Projects More capable as a writing collaborator

For DISSINET researchers: When doing strict source analysis (what do these depositions say), NotebookLM may actually be preferable to Claude Desktop for minimizing hallucination. When developing an interpretive argument or writing, Claude Projects is stronger.

Academic librarian configuring Claude Desktop for literature search

Aaron Tay — "The Agentic Researcher"

Tay configured Claude Desktop with multiple MCP servers:

Demonstrated workflow: 1. Search PubMed for papers on a topic 2. Rank top results 3. For each top result: retrieve citing and contrasting citations from Scite 4. Synthesize the citation landscape

All in one coordinated Claude conversation, with Claude orchestrating across multiple databases.

For DISSINET: The same configuration with humanities-relevant databases (JSTOR, Google Scholar, domain-specific archives) would give researchers a research assistant that actively searches rather than just analyzing pre-loaded documents.

Also from Aaron Tay — PubMed MCP enabling live search strategy testing:

Claude Desktop Projects: practical configuration for a research paper

(Based on synthesized best practices from multiple sources)

A well-configured Claude Desktop Project for an ongoing research paper:

Project instructions:

This project is for my article on [topic].
Core argument: [1-2 sentences].
Key sources: [list].
Output style: academic English, Chicago author-date citations.
When I ask for writing help, maintain my voice — do not produce generic academic prose.
When I ask for source analysis, cite page numbers.
Do not add unsolicited suggestions.

Attached files:

Result: Every conversation in the Project starts with full context. Claude knows the argument, the sources, the stage of writing. No re-explanation needed.

Claude Desktop for peer review and manuscript feedback

(Synthesized from multiple researcher accounts)

A common use pattern among academics: using Claude Desktop Projects to give structured feedback on manuscripts.

Setup:

What works: Structural feedback, identifying missing citations, flagging methodological inconsistencies, checking that the abstract matches the actual argument.

What does not work well: Evaluating the novelty of the contribution (Claude cannot reliably assess what is genuinely new in a field), assessing the quality of primary source interpretation (requires domain expertise), catching subtle disciplinary errors.

The honest assessment: Desktop's ceiling

From multiple researcher accounts, Claude Desktop reaches its ceiling when:

  1. Files change. Your draft evolves but the Project copy is stale. Re-uploading is friction.
  2. The source set grows. Adding sources 21–30 to a Project that already has 20 means managing upload limits and stale copies.
  3. You need systematic processing. "Go through 40 PDFs and extract X" is not practical in Desktop — it must be done folder by folder, file by file, with manual copy-paste.
  4. You want to build something. Any task that ends in "save this to a file" or "now process the next one the same way" hits Desktop's wall.

The ceiling is real but not immediately obvious. Researchers hit it after months of use, when their projects become large enough to expose it. Understanding the ceiling in advance helps researchers decide earlier to move to Claude Code for the appropriate tasks.

Related