Reading, Note-Making, and Writing Workflows with Claude

What this covers: How researchers integrate Claude into their core intellectual work — processing readings, building knowledge systems, and drafting academic writing.

The fundamental tension: your voice vs. Claude's

Before any workflow: the most important honest observation from researchers who have tried this.

"Without explicit guardrails, Claude will get a bit aggressive when editing my work, and its voice seeps in." — Greg Baugues

"Claude's language refinement challenges the core principle that note-taking should be purely in one's own words." — Devanshi Brower (on Zettelkasten + Claude)

For humanities and social science researchers, where intellectual ownership of argument and interpretation is foundational, this is not a minor stylistic concern. Claude's voice is a real risk. The workflows below are structured to address it — keeping Claude in an organizational and retrieval role, with synthesis and prose firmly owned by the researcher.

The key framing (from Noah Brier, knowledge management practitioner):

Use Claude in reading mode, not writing mode. Retrieval, organization, pattern-flagging — yes. Interpretive leaps and prose generation — that is your job.

1. Reading workflow: from highlights to research insight

Best entry point. No technical setup required.

Mark Carrigan's workflow (academic social theorist, writes about AI and knowledge production):

Here are my highlights and notes from [title/author].
My research focuses on [your field/topic].

Please:
1. Identify five core themes running through what I highlighted
2. Analyse my marginal notes — what do they reveal about my reactions and concerns?
3. Generate five exploratory questions to drive future inquiry in my research
  1. Optional, and particularly powerful: ask for a meta-analysis — how does this AI-assisted reading session itself exemplify themes from the source material?

What works: Pattern identification across dispersed highlights; research question generation; making implicit interests in your reading visible to yourself.

What requires human judgment: Whether the themes are genuinely significant or just frequent. Claude identifies patterns in your highlights; you evaluate whether they represent something worth pursuing.

For DISSINET: Works for any annotated reading — secondary literature, published primary source editions, digital texts. Export format does not matter; any plain text input works.

2. Note-making: Claude as organiser, not author

The key principle: Claude processes and organises your notes; it does not write them.

The Zettelkasten + Claude approach

The Zettelkasten method (atomic notes, each containing one idea, linked to related notes) is already widely practiced in humanities and social sciences. Claude can automate its most labor-intensive mechanics.

Documented workflow (synthesized from multiple practitioners):

  1. Reading phase (human): Read and make rough notes — bullets, fragments, reactions. Do not worry about format.
  2. Atomic note generation (Claude): Submit rough notes to Claude Code; ask it to create individual .md files, one idea per file, with suggested links to existing notes.
  3. Synthesis (human + Claude): Once a body of notes exists, ask Claude to identify patterns, unlinked concepts, or gaps. Human evaluates and refines.

The critical rule: Your notes must be in your words before Claude touches them. Claude's job is reorganization, not rewriting. If Claude is generating the note content, you have lost the core benefit of Zettelkasten — building understanding through the act of formulation.

I have rough reading notes in notes/raw/reading_2025_03_12.md
Please:
1. Break them into atomic ideas — one idea per file
2. Save each to notes/atomic/ with a descriptive filename
3. For each note, suggest 1-2 existing notes it might link to (list the filenames in /notes/atomic/)
Do not rewrite my words — keep my phrasing, just restructure

Obsidian + Claude Code backlinking

For researchers maintaining an Obsidian vault: Claude Code can read the entire vault and perform automated tasks that would take hours manually.

Kyle Gao's documented use case (https://kyleygao.com/blog/2025/using-claude-code-with-obsidian/):

Read my journal entry from today (notes/journal/2025-03-12.md)
Find all people, places, books, and concepts mentioned.
For each: check if a note exists in /notes/persons/, /notes/places/, /notes/concepts/
If a note exists: add a wiki-link in the journal entry
If no note exists: create a stub note and add the link

An Obsidian vault is just a folder of markdown files on your disk — exactly what Claude Code is designed to navigate. No special plugins needed.

Eleanor Konik (12-million-word vault) uses Claude Desktop + Filesystem MCP + Readwise MCP to handle organizational debt at scale: identifying inconsistent frontmatter, migrating data between note formats, detecting patterns across the vault. Her verdict: real utility for "organizational grunt work"; she would not trust Claude for "truly complex things" without debugging experience.

3. Reading workflow: PDF analysis at scale

Works in Claude Desktop (upload) or Claude Code (folder access).

Single paper

Please read this paper and give me:
1. Main argument (2 sentences)
2. Key evidence and how it supports the argument
3. Methodological approach and any limitations the author acknowledges
4. How this connects to [your research question/topic]
5. One critical question I should think about before accepting the argument

Multiple papers simultaneously

Upload or point Claude Code at a folder of PDFs:

I have 8 papers in /reading/seminar_03/
For each paper:

- Main argument (1 sentence)

- Method used

- Key finding

- How it agrees or disagrees with the others on [topic X]

Output as a comparison table in seminar_notes.md

The chapter-by-chapter approach for long texts

For books or long documents, process section by section rather than submitting the whole:

This is Chapter 3 of [book]. We are building a cumulative reading note.
Previous chapters covered: [brief summary of chapters 1-2]
For Chapter 3:
1. What new argument or evidence does it introduce?
2. How does it relate to or modify what came before?
3. What questions does it raise that subsequent chapters must answer?

Always verify: Any specific citations, statistics, or factual claims Claude includes in a summary. Hallucination rate in specialized academic content can be significant.

4. The historian's calibration: what fails for archival work

From Benjamin Breen (historian, UC Santa Cruz): https://resobscura.substack.com/p/generative-ai-for-historical-research

What AI does well for historians:

What fails — specific to historical/archival work:

Breen's overall verdict: "Augmentation, not automation." Use AI for "getting the gist" on sources you will engage with more deeply; never for precise translation or as the final word on specialized content.

5. Writing workflow: collaboration with guardrails

The right role for Claude in academic writing

From multiple researcher accounts, the most productive framing is:

What works well:

What goes wrong:

The voice-preservation prompt

When asking for editorial help, always specify:

Revise the following for clarity and flow.
Keep my argument, my examples, and my phrasing as close as possible.
Do not add new claims, do not cut substantive content, do not smooth over things I hedge deliberately.
Track what you changed and why.

The adversarial reviewer prompt

For checking your argument before submission:

You are a skeptical peer reviewer for [journal name].
Read this [section/paper].
What are the three strongest objections you would raise?
What is missing from the literature engagement?
What would you ask for in revision?
Do not be polite — I need the hard version.

Greg Baugues' dictation workflow

For generating first-draft prose without typing (useful for processing fieldwork, interview reactions, or just getting thoughts down fast):

  1. Set up Claude with your document open (Desktop + Filesystem MCP, or Claude Code)
  2. Use a voice-to-text tool (SuperWhisper on Mac; Windows Voice Typing built-in; any dictation app)
  3. Ask Claude to "interview" you — it asks questions from your outline one at a time
  4. Dictate your answers
  5. Claude inserts them into the document
  6. You do a final editing pass

The result is your words, your thinking — Claude is just the structure that pulls them out. Baugues wrote 2,000 words in 90 minutes this way.

6. Zotero integration in reading/writing workflows

Once the Zotero MCP is connected (see A10), it changes the reading and writing workflow significantly:

During reading:

I am reading about [topic]. Search my Zotero library for papers on this topic.
For the top 3 results: give me the main argument and how it relates to [specific aspect].

During writing:

I am writing the literature review section. The argument so far is: [summary].
Search my Zotero library for papers that:
1. Directly support this argument
2. Present the strongest counterarguments
For each, give me the Zotero key so I can insert the citation.

The Zotero + Obsidian + Claude triangle (documented by Alexandra Phelan, Georgetown; spektrl, thesis workflow):

This is the most mature integrated workflow found for academic researchers. Setup friction is real — but once configured, the three tools together cover the full reading-to-writing pipeline.

7. The DISSINET-specific reading pipeline

Putting it together for typical DISSINET work:

Stage 1: New source arrives (PDF scan of archival document) → Claude Code: first-pass overview, named entity extraction, save to first_pass/source_name.md

Stage 2: Close reading (human) → You read, annotate, and make rough notes

Stage 3: Note processing (Claude Code) → Convert rough notes to atomic notes, suggest links to existing notes, add to Obsidian vault

Stage 4: Synthesis (human + Claude) → "Given my notes on sources X, Y, Z, what are the main tensions and unresolved questions?" → You evaluate Claude's synthesis; you write the interpretation

Stage 5: Writing (human with Claude assistance) → Claude checks argument structure, flags gaps, helps clarify prose → You own the words

Related