paper-reading

name: paper-reading description: | Active reading and analysis of research papers. Use when reading academic papers, extracting insights, summarizing findings, or evaluating research quality. Triggers: "read this paper", "analyze paper", "summarize paper", "what does this paper say", "extract insights", "paper review", "literature review" argument-hint: "[paper-path]" agent: Explore allowed-tools: Read, Grep, Glob

Paper Reading & Analysis

Active reading strategies and analysis frameworks for research papers. Focus on extracting actionable insights efficiently.

Reading Strategy

Three-Pass Approach

Pass 1: Quick Assessment (5-10 min)

• Read title, abstract, introduction, conclusion
• Skim figures and tables
• Goal: Determine relevance and main contribution
• Output: Quick summary + relevance score

Pass 2: Selective Deep Dive (20-30 min)

• Read methodology section (if relevant to implementation)
• Examine key results and figures in detail
• Review related work for context
• Goal: Understand approach and validate claims
• Output: Technical summary + key findings

Pass 3: Critical Analysis (30-60 min, only if needed)

• Full careful reading
• Verify experimental setup
• Check reproducibility details
• Assess limitations and future work
• Goal: Deep understanding for implementation or critique
• Output: Comprehensive analysis

When to Use Each Pass

Context Management for Papers

Papers are large content sources (often 20-50+ pages). Apply context management principles to read efficiently without exhausting context. See also @.claude/skills/context-management for general strategies.

Progressive Reading Phases

Match reading depth to context budget:

Rule: Only proceed to next phase if paper is relevant. Most papers stop at Abstract phase.

Save Summaries to Files

Always save paper summaries to files rather than keeping full paper content in context:

✅ GOOD: Save summary to file
"Saved summary to docs/papers/summaries/investsuite-ivar-summary.md
Key points: [brief list]"

❌ BAD: Keep full paper in context
[Full paper content filling context...]

Summary File Location: docs/papers/summaries/{paper-slug}-summary.md

When to Save:

• After Pass 1 (Quick Assessment) → Save quick summary
• After Pass 2 (Technical Summary) → Save technical summary
• After Pass 3 (Critical Analysis) → Save comprehensive analysis
• Before reading next paper → Save current findings

Context Checkpoints

When reading multiple papers:

**Context Check**: ~60% used
- Papers read: 3 ✓
- Summaries saved: 3 ✓
- Remaining: 2 papers

**Action**: Continuing normally

**Context Check**: ⚠️ ~85% used

**Saving state**:
- Paper summaries → saved to docs/papers/summaries/
- Key insights → saved to docs/papers/insights.md

**Options**:
A) Read remaining papers with Pass 1 only (quick assessment)
B) Focus on most relevant paper (deep dive)
C) Wrap up with synthesis document

Search Before Reading

Before reading a new paper:

1. Check docs/papers/_index.md for existing entry
2. Check docs/papers/summaries/ for existing summary
3. If summary exists, read it first (saves context)
4. Only read full paper if summary is insufficient

Proactive Warnings

Before reading multiple papers:

"About to read 5 papers (~150 pages total).

**Estimated Context Impact**: ~90% usage

**Smarter Approach**:
- Read abstracts first (Pass 1) → ~20% context
- Save quick summaries to files
- You choose which papers need deep dive
- I read only selected papers fully

This preserves context. Proceed with Pass 1 only?"

Literature Review Context Strategy

When doing literature reviews (multiple papers):

1. Phase 1: Discovery (Low context)

   • Read abstracts of all papers
   • Save quick assessments to files
   • Rank by relevance

2. Phase 2: Selective Deep Dive (Medium context)

   • Read technical summaries of top 3-5 papers
   • Save summaries to files
   • Build synthesis document

3. Phase 3: Targeted Analysis (High context, only if needed)

   • Full read of 1-2 most critical papers
   • Save comprehensive analysis

Never read all papers fully in one session.

Analysis Frameworks

Quick Assessment Template

For initial evaluation (~2 min read):

## Paper: [Title]

**Metadata**
- Authors: [List]
- Year: [Publication Year]
- Venue: [Conference/Journal]
- DOI/ArXiv: [Link]

**Quick Summary**
[1-2 sentence overview of the paper's main contribution]

**Relevance Score**: [High/Medium/Low] for [current task]
**Key Insight**: [One sentence takeaway]
**Action**: [Read more / Archive / Deep dive]

Technical Summary Template

For papers requiring understanding:

## Technical Summary: [Paper Title]

### Problem Statement
[What problem does this address?]

### Core Contribution
[Main novelty or advance]

### Methodology
- **Approach**: [High-level description]
- **Key Innovation**: [What's new vs prior work]
- **Technical Details**: [Architecture/algorithm/framework]

### Experimental Setup
- **Datasets**: [What data was used]
- **Baselines**: [Compared against what]
- **Metrics**: [How success was measured]

### Key Results
- **Main Findings**: [Quantitative results with numbers]
- **Performance**: [Comparison to baselines]
- **Ablation Studies**: [What components matter]

### Practical Takeaways
- **Actionable Insights**: [What can we use?]
- **Implementation Notes**: [How to apply?]
- **Caveats**: [Limitations to consider]

Executive Summary Template

For non-technical stakeholders:

## Executive Summary: [Title]

**What They Did**: [Plain language explanation]

**Why It Matters**: [Significance and impact]

**Key Finding**: [Main result in simple terms]

**Bottom Line**: [Practical implication for our work]

Critical Analysis Template

For deep evaluation:

## Critical Analysis: [Paper Title]

### Strengths
- [What this paper does well]
- [Novel contributions]
- [Sound methodology]

### Limitations
- [Acknowledged by authors]
- [Potential issues identified]
- [Missing experiments]

### Reproducibility
- [ ] Code available?
- [ ] Data available?
- [ ] Clear enough to implement?

### Future Work
- [Authors' suggestions]
- [Open questions]
- [Research gaps]

### Relevance to Our Work
- [How does this relate?]
- [What can we adopt?]
- [What should we test?]

Active Reading Techniques

Question-Driven Reading

Before reading, ask:

1. What problem does this solve? (Introduction)
2. How is it different from prior work? (Related Work)
3. What's the core idea? (Methodology)
4. Does it actually work? (Results)
5. What are the limitations? (Discussion/Conclusion)

Extract-As-You-Read

While reading, capture:

• Key definitions: Technical terms, metrics, concepts
• Formulas/algorithms: Mathematical expressions, pseudocode
• Numbers: Performance metrics, dataset sizes, hyperparameters
• Citations: Important related papers to follow up
• Open questions: What's not addressed?

Cross-Reference Strategy

When reading multiple papers:

1. Build concept map: How papers relate to each other
2. Compare approaches: Different solutions to same problem
3. Track evolution: How ideas developed over time
4. Identify consensus: What do most papers agree on?

Quality Assessment

Credibility Checklist

• Venue: Top-tier conference/journal?
• Authors: Recognized researchers?
• Reproducibility: Code/data available?
• Methodology: Sound experimental design?
• Results: Statistically significant?
• Peer Review: Rigorous venue?

Red Flags

⚠️ Warning signs:

• Unrealistic claims without evidence
• Missing experimental details
• No comparison to baselines
• Cherry-picked results
• Lack of ablation studies
• Poor writing/clarity (often indicates poor thinking)

When to Trust Results

✅ High confidence:

• Reproducible code available
• Multiple independent validations
• Consistent with related work
• Clear methodology
• Honest about limitations

Literature Review Synthesis

For comparing multiple papers:

## Literature Review: [Topic]

**Papers Analyzed**: [Count]
**Date Range**: [Years covered]

### Evolution of Ideas
1. **Early Work ([years])**
   - Focus: [approach]
   - Key papers: [List]
   - Limitations: [Issues]

2. **Current State ([years])**
   - Refined techniques: [Modern approaches]
   - SOTA results: [Best performance]
   - Key papers: [List]

### Consensus Findings
- [Agreement across papers]
- [Well-established techniques]
- [Proven approaches]

### Controversial Areas
- [Disagreements]
- [Conflicting results]
- [Open debates]

### Research Gaps
- [Under-explored areas]
- [Future directions]
- [Opportunities]

Integration with Project

When Reading Papers for This Project

1. Check existing knowledge: Look in docs/papers/_index.md first
2. Extract actionable insights: Focus on what can be implemented
3. Link to codebase: Note where concepts are used (e.g., "iVaR from InvestSuite paper")
4. Update index: Add new papers to docs/papers/_index.md with brief description
5. Document connections: Link papers to research findings in docs/research/

Paper Storage

• Location: docs/papers/ (flat structure)
• Index: docs/papers/_index.md (maintain with short descriptions)
• References: Use relative paths in documentation (e.g., docs/papers/paper-name.pdf)

Remember

Focus on extracting actionable insights rather than comprehensive summaries. Match the analysis depth to the paper's relevance to your current task. Use the three-pass strategy to efficiently triage papers, then deep-dive only when needed.

Key principle: Read to understand, not to archive. Extract what matters for the current work, document connections to the codebase, and move on.