Writing Academic Papers with Claude + Obsidian

What This Guide Is

This guide provides a workflow for AI-assisted academic writing. It covers the prose-heavy parts of scholarship: literature synthesis, argumentation, drafting, and revision.

You will use Obsidian as your knowledge base and Claude as your writing partner. The web interface works for the full workflow; Claude Code adds the ability to read and write files in your vault directly. Despite its name, Claude Code is not just for programmers. It excels at any task involving structured text and files.

The Workflow Overview

This accelerates: literature synthesis, drafting, structuring arguments, iterative revision.

This does not replace: your domain expertise—running experiments, conducting interviews, close reading, or formal derivation.

The Core Insight

AI output quality depends almost entirely on input context quality.

Dump a vague request → get hollow, generic text.

Provide structured, compressed context → get useful output.

This is Context Engineering: actively managing what the AI sees. The documents you create aren't just notes. They shape what AI can do for you.

What is a Vault?

A "vault" is Obsidian's term for a folder of interconnected markdown files. Unlike traditional note-taking apps, Obsidian stores everything as plain text files on your computer. No proprietary formats, no lock-in.

Why this matters for AI-assisted writing

Vault Structure

Create this folder structure in your Obsidian vault:

/Academic-Vault/
├── CLAUDE.md                    # Your AI constitution
├── me/                          # Your intellectual identity
│   ├── style.md
│   ├── positions.md
│   └── challenges.md
├── References/                  # All sources, linked across projects
├── Attachments/                 # PDFs, data files
└── projects/
    └── Paper-Name/
        ├── CLAUDE.md            # Project-specific rules
        ├── context/
        │   ├── research.md
        │   ├── data.md
        │   ├── requirements.md
        │   └── design.md
        ├── drafts/
        ├── feedback/
        └── journal.md

Why This Structure

• /me carries your intellectual identity across projects: your writing voice, preferred methods, and theoretical commitments
• /References accumulates over time. Sources link across papers, building a personal knowledge graph.
• /projects/Paper-Name/context/ holds the distilled context AI actually uses (not raw materials, but compressed insights)
• journal.md documents your process for future reuse and methodological transparency

Safety Setup

Before granting Claude Code file access to your vault:

1. Set up real backups.

   Obsidian's File Recovery is a snapshot tool, not a true backup. For agentic workflows where AI can write files, use Git, Time Machine, or maintain an archive folder that the AI has no write access to. If something goes wrong, you need recovery outside the system the AI can touch.

2. Understand the access model.

   Claude Code can read and write files in any folder you give it access to. This is powerful but requires trust. Real backups ensure that trust is recoverable.

3. Check your target journal's AI policy.

   Some publishers (Nature, Science, Elsevier, and others) prohibit or restrict uploading unpublished manuscripts to generative AI. Verify compliance before using this workflow on work intended for specific venues. Never use this workflow for confidential peer review materials.

Phase 1: Preparation

The purpose of this phase is collection, not analysis. Resist the urge to structure too early. You don't yet know what will matter most.

1. Create your project folder in your Obsidian vault under /projects/Paper-Name/ with the subdirectories shown above (context/, drafts/, feedback/)
2. Write your initial thesis and questions in context/requirements.md. This doesn't need to be polished. Just capture what you're trying to argue and what you need to figure out.
3. Collect sources: Download PDFs to /Attachments, create notes in /References. How you structure these notes is up to you. Some prefer detailed annotations, others just bibliographic info and a few key quotes. The point is having sources accessible and linkable.
4. Start journal.md with today's date and your initial state: What do you already know? What are you uncertain about? What's your current best guess at the argument?

Don't organize yet. Just collect.

Phase 2: Exploration & Mapping

This phase produces working ideas, not polished documents. You're having a conversation with Claude to stress-test your thinking before committing to a structure.

Questions to Explore

• What's actually possible with my sources/data?
• What are the strongest objections to my argument?
• What adjacent fields have relevant frameworks I'm missing?
• Where does my thesis need refinement?

Key Prompt Pattern

I'm developing a paper arguing [THESIS].

Before I commit:
1. Steel-man the strongest opposing view
2. Identify the 3 weakest points in my argument
3. What evidence would change your assessment?

Be critical, not agreeable.

Document in journal.md

• What you tested
• What held up, what didn't
• Dead ends and why
• Sources to verify

Phase 3: Distillation

Why this matters: AI performance degrades with context length ("context rot"). More tokens ≠ better output. You want maximum information in minimum space.

The Context Documents

The Craft of Distillation

Bad distillation = dumping everything potentially relevant

Good distillation = only what's necessary for THIS paper, compressed to essentials

Ask yourself:

• If I cut this, would the AI output suffer?
• Am I describing or just copying?
• Is this already in AI's training data? (Don't explain what regression is, or who Foucault was.)

Verification before finalizing: Before you consider these documents complete, verify that key sources actually say what you think they say. Misunderstandings from Phase 2 will propagate through the entire draft if you don't catch them here.

Phase 4: Implementation

Step 1: Generate First Draft

Provide your context documents to Claude. Request a complete draft.

Write a complete draft based on this context:

[Paste or reference research.md]
[Paste or reference requirements.md]
[Paste or reference design.md]

Requirements:
- Every claim grounded in provided sources
- No filler phrases
- Flag uncertainties explicitly

Step 2: Expect It to Be Bad

First drafts are typically hollow: structurally correct but thin on substance. This is normal. The draft gives you something concrete to push against.

Step 3: Push Back with Specificity

Vague critique → vague revision.

Specific critique → actual improvement.

Problems with this draft:

1. Paragraph 3 claims "scholars have debated X" without specifics.
   Name the scholars, state the positions, cite the sources.

2. The transition from section 2 to 3 is asserted, not argued.
   What's the logical connection?

3. "It is important to note" appears 4 times.
   Cut all instances. Make direct claims.

4. You cite Martinez 2021 but this isn't in my sources.
   Remove or replace with provided sources only.

Revise addressing these specifically.

Step 4: Peer Review Cycles

Use a fresh session (new conversation) as critical reviewer. The session that wrote the draft has seen all your context and reasoning. It's primed to defend its own output rather than critique it. A fresh session approaches the text without that history.

You are a critical peer reviewer. Evaluate this draft:

[Paste draft]

Assess:
- Argument structure and logical gaps
- Evidence quality and support for claims
- Actual contribution vs. restatement
- Weaknesses a skeptical reviewer would attack

Be harsh. I need criticism, not encouragement.

Save feedback in /feedback → feed back to writing session → iterate 2-3 times.

Step 5: Human Editing

Take over. This is where your intellectual fingerprint goes.

Questions to ask yourself:

• Would I defend this claim in a seminar?
• Is this MY argument or generic synthesis?
• Can I explain this without notes?

Step 6: Source Verification

Check every citation against actual sources. AI generates plausible-sounding but wrong references confidently. You should have caught major issues in earlier verification checkpoints, but do a final systematic check.

Create feedback/verification-log.md:

## [Source]
- Claim in draft: "..."
- Actual text: "..." (p. X)
- Status: ✓ Verified / ✗ Remove / ~ Revise

Domain Adaptation

The workflow is universal. What changes is what fills the context documents.

Effective Critique

The skill that most improves AI output is specific, substantive pushback.

Domain-Specific Examples

• Quantitative: "You describe 'significant results' without specifying significance level, test used, or effect size."
• Experimental: "Methods must follow CONSORT structure. Reorganize into: participants, interventions, outcomes."
• Theoretical: "The proof sketch in section 3 skips from step 2 to step 4. Make the intermediate step explicit."
• Qualitative: "This reads as naive realism. Add reflexivity statement addressing researcher positionality."

The Journal

journal.md serves three functions:

1. Process documentation: What you did, what worked, what didn't
2. Knowledge capture: Insights that don't fit in context docs but matter
3. Future reference: Reusable patterns for similar projects

Entry Format

## YYYY-MM-DD

### Did
- [Actions taken]

### Learned
- [Insights, surprises, dead ends]

### Next
- [What follows from today]

Final Entry

## YYYY-MM-DD (Complete)

### What worked
### What I'd change
### Time spent
### Reusable for future

Common Problems

AI keeps producing generic output. Your context is too vague or too long. Distill harder. Be more specific in requirements.

AI ignores my theoretical framework. State it explicitly in prompts: "My approach is [X]. Do not use [Y] framing."

AI hallucinates citations. Constrain: "Only cite from these sources: [list]." Verify everything anyway.

Output improves then degrades. Context window filling up. Start fresh session with just context documents.

Workflow feels slow. For shorter papers, combine documents. Skip peer review cycles. The full workflow is for substantial research papers.

Quick Reference

The Four Phases

1. Preparation — Gather materials, initial questions, start journal
2. Exploration — Test ideas with AI, map possibilities, verify critical claims
3. Distillation — Compress into context documents, verify sources
4. Implementation — Draft → Push back → Review cycles → Verify → Human editing

Context Documents

• research.md — Framework + sources
• data.md — What you're working with
• requirements.md — Your questions
• design.md — Structure + venue

Core Principles

• Your thinking first
• Compress, don't dump
• Critique specifically
• Verify throughout (not just at the end)
• Journal the process

Sources

• Christopher Pollin: "Promptotyping". L.I.S.A. Wissenschaftsportal, 2026.
• Andrew Maynard: "I cracked and wrote an academic paper using AI". The Future of Being Human, 2026.
• Steph Ango: "How I use Obsidian". stephango.com.
• Ethan Mollick: Co-Intelligence: Living and Working with AI. Portfolio/Penguin Random House, 2024.
• The Neuron: "How to Turn Claude Code Into Your Personal AI Assistant". 2025.

Note on Tools

This guide is written for Claude Code and Obsidian as they exist in early 2026. Tools evolve. New AI assistants will emerge, interfaces will change, capabilities will expand.

The methodology is more durable:

• Context engineering (compress, don't dump)
• Iterative critique (push back with specificity)
• Verification throughout (don't trust, verify)
• Process documentation (journal your work)

If you're reading this with different tools, the principles still apply. Adapt the specific instructions to whatever AI and knowledge management system you're using.