Always fetch the URL first using WebFetch to inspect the response
Check whether the response is PDF binary data or readable HTML/text
Signs of a PDF response: binary/garbled content, mentions of "PDF binary data", application/pdf content type, or the tool saving a .pdf file locally
Signs of an HTML response: readable structured text, title/abstract visible, normal markdown output
If the fetched content is a PDF (binary, unreadable):
Read the locally saved PDF file using the Read tool on the path reported in the WebFetch result (e.g., /Users/.../.pdf)
For large PDFs (> 10 pages), read in chunks using the pages parameter (e.g., pages: "1-10", then "11-20", etc.)
DO NOT perform any web search or look for an arXiv version
Use the original URL as the url frontmatter field in the paper note
If the fetched content is readable HTML:
Extract all paper content directly from the WebFetch response
Use the URL as the url frontmatter field
When given a non-URL input (arXiv ID, paper title, DOI):
ALWAYS try the arXiv HTML version first: https://arxiv.org/html/[paper-id]
If HTML is unavailable (404, missing, or incomplete), download the TeX source: https://arxiv.org/src/[paper-id]
The source is a .tar.gz archive — extract it and read the .tex files
Focus on the main .tex file (often main.tex or the largest .tex file)
Extract content from the TeX source as you would from HTML
Only use PDFs as a last resort if neither HTML nor TeX source is available
Workflow
0. Read the template FIRST
MANDATORY FIRST STEP:
Use view tool on: Templates/Paper Note.md
Study the structure before creating the paper note.
1. Create paper note
File name: exact paper title
if present in paper title, exchange : for -
Location: Reading/
Copy template structure EXACTLY from Templates/Paper Note.md
Fill each section with bullet points under the subsection prompts
Do NOT remove subsection headings or prompts
Link to existing projects immediately in "Linked Research Projects"
Only propose new projects - don't create them yet
If the paper itself mentions a code repository (e.g., a GitHub link in the abstract, introduction, or footnotes), add that URL to the code property in the frontmatter; otherwise leave it empty — do NOT search online for a code repo
2. Fill template sections
Answer each subsection prompt with bullet points:
🌍 Background: Answer both subsection prompts
❓ Problem Statement: Answer all three subsection prompts
💡 Key Idea: Answer all three subsection prompts
🧭 High-Level Approach: Describe method/framework
🔬 Technical Details / Analysis: Answer all three subsection prompts
📊 Main Results / Observations: List findings
🧩 Interpretation: Answer both subsection prompts
⚠️ Limitations / Open Questions: List items
🔗 Connections to Other Work: Fill three subsections with [[Paper Title]] wiki links
Link to ALL influential papers mentioned or referenced in the current paper
Link even if paper notes don't exist in vault yet (they will be added when those papers are read)
MUST use exact, full paper titles in Title Case
MUST format as wiki links: [[Title]] not plain text
Extract titles from paper references/citations when available
🚀 Relevance for My Research: Answer two prompts + add project links
CRITICAL: Write technical details as plain text. Do NOT create wiki links for concepts, methods, or techniques.
3. Identify existing projects and propose new ones
Part A: Link to existing projects immediately
Search for existing projects in Projects/ that relate to the paper
Add links to these existing projects in the paper note under "Linked Research Projects"
No user approval needed for linking to existing projects
Note: Reciprocal links (project → paper) will be added in step 4
Part B: Propose new projects (requires user approval)
CRITICAL: Ground proposals in the paper itself, not inferred extensions
Extract ideas directly from the paper's content:
Explicit open questions the paper identifies (e.g., "future work", "unclear", "remains to be seen")
Limitations the authors acknowledge
Unexplained observations or surprising results mentioned in the paper
Assumptions stated but not tested
Gaps the paper identifies in prior work
Follow-up experiments the paper suggests
DO NOT infer extensions beyond what the paper discusses. If the paper doesn't mention it, don't propose it.
Before creating ANY new project files:
Re-read the paper's "Limitations / Open Questions" section you filled in step 2
For each open question or limitation that could be a standalone project:
Proposed project title (describes the idea, grounded in paper content)
Quote or paraphrase from the paper showing where this question/gap appears
1-2 sentence description of what it would explore
Note if it's related to any existing project
Use AskUserQuestion tool to let the user select which projects to create:
Use multiSelect: true so the user can pick multiple projects
Each option: label = project title, description = 1-line summary of what it explores
Include a "None" option with description "Don't create any new projects"
Example:
question: "Which projects should I create from this paper?"
header: "Projects"
multiSelect: true
options:
- label: "Project Title A", description: "Explores X based on paper's open question about Y"
- label: "Project Title B", description: "Tests whether Z holds, motivated by limitation W"
- label: "None", description: "Don't create any new projects"
Wait for user selection
Only create the selected projects (skip if "None" selected)
Novelty and feasibility assessment (for each proposed project):
novelty (1–10): How novel is the idea relative to existing work? (1 = incremental, 10 = highly original)
feasibility (1–10): How feasible is it to execute with current methods/resources? (1 = very difficult, 10 = straightforward)
Assess both scores based on the paper's content and context; include them in the description shown to the user
Project title examples:
Detecting Representation Drift During Fine-Tuning
Preventing Emergent Misalignment via Gradient Monitoring
For each approved new project:
create a new file in Projects/
use Templates/Project.md as template
status: seed
fill novelty and feasibility frontmatter fields with assessed 1–10 scores
add 1–3 bullet points minimum
link the current paper under Relevant Papers
4. Create approved new projects and finalize bidirectional links
After receiving user approval, create the new project files:
Create only the approved new projects in Projects/
Use Templates/Project.md as template
Fill with the descriptions you prepared
Link the current paper under Relevant Papers in each project
CRITICAL: Bidirectional linking for existing projects
For every project linked from the paper note (both existing and newly created):
Open the project note in Projects/
Locate the "Relevant Papers" section
Add a wiki link to the paper note if it's not already there:
Format: [[Paper Title]] (the paper note file name)
Use Title Case with the exact paper title
Place it in the "Relevant Papers" section
Save the project note
This ensures the graph remains consistent: if paper → project link exists, then project → paper link must also exist.
Then update the paper note:
Add newly created project links to "Linked Research Projects" (existing projects should already be linked from step 1)
Ensure paper links in "🔗 Connections to Other Work" subsections are complete
CRITICAL: Use exact paper title in Title Case when linking to papers
CRITICAL: Format paper references as wiki links: [[Paper Title]] not plain text
If you don't know the exact title, use a descriptive title in Title Case
Examples:
✅ [[Sparse Autoencoders Find Highly Interpretable Features in Language Models]]
✅ [[Concept Bottleneck Models]]
❌ "Bricken et al. 2023" (not a wiki link)
❌ [[sparse autoencoders]] (not title case, too generic)
VERIFY: Every [[link]] in the paper note is either:
A file in Projects/ (project note) - appears under "Linked Research Projects"
A paper title as wiki link - appears under "Connections to Other Work" as [[Title Case Paper Title]]
Paper notes may not exist in vault yet, but will be added when those papers are read
Use exact, full paper titles from references
5. Minimal quality check
Before finishing:
ALL template sections present with emojis
ALL subsection prompts (### questions) present
Paper note contains ≥ 5 internal links (ONLY to projects/papers)
At least one project note created or reused
No paragraph longer than 3 bullets
VERIFY: Every wiki link is to a file in Projects/ or Reading/
No concept links (attention, transformers, etc.) - only plain text
6. Update maintenance counters
After completing the paper note (and any new projects):
Read .claude/vault_maintanance.md
Increment papers_since_audit_links, papers_since_vault_index, and papers_since_validate_links each by 1
If any new project files were created, increment projects_since_audit_links and projects_since_vault_index each by the number of new projects created
Write the updated values back to .claude/vault_maintanance.md
Output Style
bullet points only
fragments preferred
dense information
no narrative text
preserve ALL template structure, emojis, and subsection prompts
technical terms as plain text, NOT wiki links
links ONLY to projects (Projects/) or papers (Reading/)
don't include : in file names. Exchange : always with -
Goal
Convert reading into:
explicit mechanisms
reusable research ideas
connected project graph
long-term synthesis potential
properly structured notes following template exactly
Important workflow:
Link to existing projects immediately without asking
Propose new projects and get user approval before creating them
A paper note without linked projects is considered incomplete.
A paper note that doesn't follow the template structure is considered invalid.
Self-Check Before Completing
Template structure:
Did I read Templates/Paper Note.md first? ✅
Are ALL emoji section headers present? ✅
Are ALL subsection prompts (### questions) present? ✅
Did I answer each subsection separately? ✅
Links:
For EVERY [[link]] in the paper note:
Is this link pointing to a file in Projects/? ✅
Is this link pointing to a file in Reading/? ✅
Is this link a concept/method/technique? ❌ REMOVE
For paper links: Is it in Title Case with wiki link brackets? ✅
For paper links: Did I use the full paper title, not "Author et al."? ✅
Bidirectional linking verification:
For each project linked in the paper note:
Did I add/verify the reciprocal link from project → paper? ✅
Does the project's "Relevant Papers" section include this paper? ✅
All paper-project relationships are now bidirectional ✅
If uncertain whether something should be a link: Don't link it. Use plain text.
New projects:
Did I fill novelty (1–10) in each new project's frontmatter? ✅
Did I fill feasibility (1–10) in each new project's frontmatter? ✅
Maintenance counters:
Did I increment papers_since_audit_links, papers_since_vault_index, papers_since_validate_links in .claude/vault_maintanance.md? ✅
Did I increment projects_since_audit_links and projects_since_vault_index for each new project created? ✅
