read-paper

📄 Paper Reading Skill

Read a research paper and convert it into:

• one structured paper note
• one or more linked research project notes

This skill transforms reading into reusable research ideas.

---

CRITICAL RULES (NON-NEGOTIABLE)

• Paper notes live only in Reading/
• Use Templates/Paper Note.md exactly - preserve ALL headings, emojis, and structure
• Project notes live only in Projects/
• Never create topic notes or concept-only notes
• Research ideas always live as project notes
• Paper notes must never link to MOCs
• Paper notes may link only to:
  • project notes (in Projects/)
  • other paper notes (in Reading/)
• NO OTHER LINK TYPES ALLOWED - no concepts, no methods, no techniques
• Bullet points only — no prose
• Aggressive [[wiki linking]] required BUT ONLY to projects/papers

---

TEMPLATE COMPLIANCE (MANDATORY)

BEFORE starting, read the template:

1. Use the view tool to read Templates/Paper Note.md
2. Copy the EXACT structure including:
   • All emoji section headers (🧾, 🌍, ❓, 💡, etc.)
   • All subsection prompts (### questions)
   • All markdown formatting
3. Fill in content under each subsection prompt
4. Do NOT remove, rename, or flatten any sections

Template location: Templates/Paper Note.md

Rules:

• Keep ALL emojis exactly as shown in template
• Keep ALL section headings exactly as shown in template
• Keep ALL subsection prompts (### questions) exactly as shown
• Answer each subsection with bullet points below the prompt
• Do NOT remove or rename sections
• Do NOT flatten the structure
• Place project links ONLY in "Linked Research Projects" subsection
• Place paper links ONLY in "Connections to Other Work" subsections

---

LINKING RULES (CRITICAL)

✅ ALLOWED Links in Paper Notes

1. [[Project Name]] - links to files in Projects/
2. [[Paper Title]] - links to influential papers
   • Link to papers even if they don't exist in vault yet (paper notes can be created later)
   • MUST use exact, full paper titles in Title Case
   • MUST format as wiki links with double brackets
   • Example: [[Attention Is All You Need]] not "Attention is all you need"
   • Example: [[BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding]]
   • Example of future paper link: [[Representation Engineering - A Top-Down Approach to AI Transparency]]

❌ FORBIDDEN Links in Paper Notes

• [[Transformer Architecture]] - this is a concept, NOT allowed
• [[Attention Mechanism]] - this is a technique, NOT allowed
• [[Self-Supervised Learning]] - this is a method, NOT allowed
• [[BERT]] - this is a model name, NOT allowed (unless it's a paper title)
• [[Cross-Entropy Loss]] - this is a concept, NOT allowed
• Any link that is not explicitly a project or paper

How to Handle Technical Terms

Do NOT link them. Just write them as plain text.

Examples:

• ❌ "The paper uses [[attention mechanisms]]"
• ✅ "The paper uses attention mechanisms"
• ❌ "They employ [[contrastive learning]]"
• ✅ "They employ contrastive learning"

Exception: Only link if you're creating/referencing a PROJECT about that mechanism:

• ✅ "Uses attention mechanisms → [[Scaling Laws for Attention Heads]]" (project)

---

Paper Access Rules

Step 0: Determine what the URL points to

• 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):

1. Read the locally saved PDF file using the Read tool on the path reported in the WebFetch result (e.g., /Users/.../.pdf)
2. For large PDFs (> 10 pages), read in chunks using the pages parameter (e.g., pages: "1-10", then "11-20", etc.)
3. DO NOT perform any web search or look for an arXiv version
4. Use the original URL as the url frontmatter field in the paper note

If the fetched content is readable HTML:

1. Extract all paper content directly from the WebFetch response
2. Use the URL as the url frontmatter field

When given a non-URL input (arXiv ID, paper title, DOI):

1. ALWAYS try the arXiv HTML version first: https://arxiv.org/html/[paper-id]
2. 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
3. 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:

1. Re-read the paper's "Limitations / Open Questions" section you filled in step 2
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
3. 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"
     

4. Wait for user selection
5. 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):

1. Open the project note in Projects/
2. Locate the "Relevant Papers" section
3. 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
4. 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:

1. A file in Projects/ (project note) - appears under "Linked Research Projects"
2. 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):

1. Read .claude/vault_maintanance.md
2. Increment papers_since_audit_links, papers_since_vault_index, and papers_since_validate_links each by 1
3. 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
4. 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:

1. Did I read Templates/Paper Note.md first? ✅
2. Are ALL emoji section headers present? ✅
3. Are ALL subsection prompts (### questions) present? ✅
4. Did I answer each subsection separately? ✅

Links: For EVERY [[link]] in the paper note:

1. Is this link pointing to a file in Projects/? ✅
2. Is this link pointing to a file in Reading/? ✅
3. Is this link a concept/method/technique? ❌ REMOVE
4. For paper links: Is it in Title Case with wiki link brackets? ✅
5. For paper links: Did I use the full paper title, not "Author et al."? ✅

Bidirectional linking verification:

1. 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? ✅
2. All paper-project relationships are now bidirectional ✅

If uncertain whether something should be a link: Don't link it. Use plain text.

New projects:

1. Did I fill novelty (1–10) in each new project's frontmatter? ✅
2. Did I fill feasibility (1–10) in each new project's frontmatter? ✅

Maintenance counters:

1. Did I increment papers_since_audit_links, papers_since_vault_index, papers_since_validate_links in .claude/vault_maintanance.md? ✅
2. Did I increment projects_since_audit_links and projects_since_vault_index for each new project created? ✅