paper-xray

name	paper-xray
description	Academic paper deep analysis and research automation. Four modes: 1. Single: Given a paper, perform deep structural analysis 2. Batch: Given multiple papers, parallel processing with cross-paper insights 3. Triage: Quick relevance assessment against research direction, no full notes 4. Topic: Given a research topic, auto-search, filter, analyze each paper, generate survey report Auto-archives to workspace with bidirectional links and index maintenance.
allowed-tools	Bash, Read, Write, Edit, Glob, Grep, WebSearch, WebFetch, Agent
user-invocable	true

Paper X-Ray: Paper Analysis & Topic Research

Four modes: Single deep analysis / Batch parallel / Quick triage / Topic auto-survey.

Configuration

Customize these paths before first use. All paths are relative to your workspace root (current working directory).

Variable	Description	Default
NOTES_DIR	Paper notes directory	papers/notes/
READING_LIST	Reading list file path	papers/reading-list.md
INDEX_FILE	Paper notes index file	papers/notes/index.md
COGNITIVE_BASELINE	Cognitive baseline files for personal reflection (optional)	Not enabled
RESEARCH_FOCUS	Default research direction for triage mode	Must be provided by user

Usage

# Single mode
/paper-xray <paper name | arxiv ID | DOI | URL | PDF path>

# Batch mode
/paper-xray batch: <id1> <id2> <id3> ... [category: <topic>]

# Triage mode (quick assessment, no notes generated)
/paper-xray triage: <id1> <id2> ... [context: <research direction>]

# Topic research mode
/paper-xray topic: <research topic> [scope: <constraints>] [n: <paper count>]

Single mode examples:

• /paper-xray Attention Is All You Need
• /paper-xray 2301.07041
• /paper-xray https://arxiv.org/abs/2301.07041
• /paper-xray ~/Downloads/paper.pdf

Batch mode examples:

• /paper-xray batch: 2603.04756 2504.05738 2603.05344 category: Agent

Triage mode examples:

• /paper-xray triage: 2603.04756 2504.05738 context: AI-assisted software engineering

Topic mode examples:

• /paper-xray topic: Agent in Software Engineering
• /paper-xray topic: Code Generation Evaluation Benchmarks scope: 2024-2026 n: 8

Mode Routing

Auto-detect mode based on input:

• Contains batch: prefix → Batch mode
• Contains triage: prefix → Triage mode
• Contains topic: prefix → Topic research mode
• Matches arxiv ID / DOI / URL / PDF path / specific paper name → Single mode
• Ambiguous (e.g., "latest LLM Agent progress") → Default to Topic mode

Paper Type Detection

After fetching a paper, auto-detect type and use corresponding template:

Type	Identification	Template
Research Paper	Has clear Method + Experiments, proposes new method/framework	references/template.md
Technical Report	Title contains model name (GPT/Claude/Qwen etc.), architecture+training+eval focused, no controlled experiments	references/tech-report-template.md
Survey Paper	Title contains Survey/Review/Taxonomy, primarily categorization	references/survey-analysis-template.md

Priority: Technical Report > Survey > Research Paper (default).

Principles

• Objectivity first: The analysis subject is the paper itself, not personal opinions. Deconstruct the paper first, then discuss personal insights.
• Precise terminology: Target audience is researchers/engineers. Use proper terms, but define key concepts clearly.
• Honest uncertainty: Distinguish "experimentally proven" from "author speculation". Mark insufficient information, don't fabricate.
• Structured, not narrative: Use tables, bullet points, comparisons. No essays.

---

Single Mode Execution Flow

Step 1: Fetch Paper

Route based on input type:

Input Format	Recognition Rule	Fetch Strategy
2301.07041 or arxiv:2301.07041	Digits+dot or arxiv: prefix	→ https://arxiv.org/html/{id} WebFetch
https://arxiv.org/abs/...	arxiv abs URL	→ Replace with /html/ WebFetch
https://arxiv.org/pdf/...	arxiv pdf URL	→ Replace with /html/ WebFetch
https://arxiv.org/html/...	arxiv html URL	→ Direct WebFetch
10.xxxx/...	DOI format	→ WebSearch for arxiv/html version
~/xxx.pdf or /xxx.pdf	Local path	→ Read tool to read PDF
Other URL	Generic URL	→ WebFetch
Paper name (natural language)	No match above	→ WebSearch, prefer arxiv html

Fetch priority: arxiv HTML > other HTML > PDF.

Robust fetch strategy (arxiv papers):

1. Try https://arxiv.org/html/{id}
2. If HTML unavailable (404 or incomplete), try https://arxiv.org/abs/{id} for abstract page
3. If abs page also fails, WebSearch arxiv {id} for alternative sources
4. Last resort: WebSearch paper title for any available version

Long paper handling: Fetch Abstract + Introduction + Conclusion first for global understanding, then fetch Method, Experiments etc. as needed. Don't dump everything at once.

Metadata extraction: Title, authors, publication date, venue, arxiv ID (if available).

arxiv ID year parsing: ID format YYMM.NNNNN, first two digits = year. 2602.09540 = Feb 2026, 2301.07041 = Jan 2023.

Step 2: Problem & Motivation Analysis

• Problem definition: 1-2 sentences precisely describing the core problem
• Existing method limitations: Where previous work got stuck — theoretical/engineering/data bottleneck?
• Key Insight: What did the authors see that others missed?

Step 3: Technical Approach Deconstruction

• Overall architecture: Mermaid diagram or structured description of the system/method flow
• Key technical contributions: Break down 2-3 core contributions, each explaining: what it is, why, and how it fundamentally differs from existing methods
• Key formulas/algorithms: Core formulas + symbol meanings + intuition. Don't skip the math, but make it understandable
• Design choices & trade-offs: Where did the authors make trade-offs?

Step 4: Experiments & Results Evaluation

• Experimental setup: Datasets, baselines, evaluation metrics
• Main results: Key numbers in tables, annotate the most important comparisons
• Ablation studies: Contribution of each component
• Result credibility: Statistical significance, conclusion boundaries, reproducibility (code/data/hyperparameters)

Step 5: Contributions & Limitations

• Core contributions: Substantive advancement to the field (not just restating the abstract)
• Limitations: Author-acknowledged + unmentioned but present
• Open questions: What remains unsolved?

Step 6: Field Positioning & Connections

• Academic lineage: Who does it build on? Who does it complement/contradict?
• Impact assessment: Opens new direction / improves existing method / provides new evidence
• Recommended reading: 2-3 papers most worth reading next + reasons

Step 6.5: Deep Reading — What the Paper Doesn't Say

• Implicit assumptions: Unstated assumptions — what happens if they don't hold?
• Deliberately avoided questions: Obvious comparisons/scenarios not discussed
• Stories behind the data: Anomalous numbers in results
• Deeper perspective: What does the methodology choice reveal about the authors' understanding?
• If I were the author's next step: Most valuable thing the authors didn't do

Step 6.6: Reviewer Assessment

Structured review from a top-venue reviewer perspective:

• Overall rating: Strong Accept / Weak Accept / Borderline / Weak Reject / Strong Reject
• Strengths (3-5 points)
• Weaknesses (3-5 points, each with: issue, why it matters, how to improve)
• Questions to Authors: 2-3 questions targeting the weakest points
• Missing References
• Verdict: One sentence summary

Reviews should be fair but sharp. Good papers have weaknesses, bad papers have strengths. No pleasantries.

Note: Technical reports (model releases) skip this step, use competitive analysis from the tech-report template instead.

Step 6.7: Insights & Action Items

• Directly reusable: Methods, datasets, evaluation frameworks, code
• Borrowable ideas: Problem framing, experimental design, narrative structure
• Triggered new ideas: New research questions or engineering ideas
• Action items: Specific next steps

Step 7 (Conditional): Personal Cognitive Reflection

This step is optional and only executes when configured. Users can enable it by setting COGNITIVE_BASELINE in the Configuration section to point to their personal reflection files.

If COGNITIVE_BASELINE is configured:

1. Read the specified cognitive baseline files
2. Find collision points: paper findings vs personal beliefs → reinforce/revise/overturn?
3. Output a one-sentence delta. delta ≈ 0 is a normal result — don't force it.

If not configured, skip this step entirely.

Step 8: Archive to Workspace

8.1 Determine Topic Category

Categorize paper into topic directory: {NOTES_DIR}/{topic}/

Common topics: Agent / LLM / Knowledge Representation / Software Engineering / Multimodal / RL / AI Safety. Create new topics as needed.

8.2 Generate Report File

Select template based on paper type (template.md / tech-report-template.md / survey-analysis-template.md).

File naming: {Paper Title}.md, use concise recognizable English. Save to {NOTES_DIR}/{topic}/.

8.3 Auto Knowledge Network Links

1. Search existing notes in the workspace for related papers using keyword matching
2. Add [[Related Paper]] links in the new report's "Knowledge Network" section + relationship description
3. Reverse update: In existing related paper notes, append a link back to the new paper

8.4 Update Index

• Update {NOTES_DIR}/{topic}/_index.md (create if not exists)
• Update {INDEX_FILE} under the corresponding topic category

8.5 Update Reading List

Check {READING_LIST}: if the paper has an entry, mark [ ] → [x] and append → [[{Paper Title}]].

8.6 Custom Post-Processing (Optional)

If a post-archive.sh script exists in the skill directory, execute it after archiving. This allows users to integrate custom index builders or other automation.

---

Batch Mode Execution Flow

Enters this mode when input contains batch: prefix.

B1: Parse Input

Extract paper list (arxiv IDs / URLs / names) from after batch:, plus optional category: topic.

If no category specified, auto-determine topic from paper content.

B2: Parallel Analysis

Launch independent Agent per paper, executing Single mode steps 1 through 6.7 (steps 2-6.7 for analysis, step 8 for archiving).

Each Agent's prompt template:

You are a paper analysis expert. Perform deep structural analysis on the following paper.

Paper: {arxiv ID or title}
Fetch method: WebFetch https://arxiv.org/html/{id} (if fails, try abs page or WebSearch)

Execute Single mode steps 1 through 6.7 completely, output structured report.

Paper type detection: Determine if research paper/technical report/survey based on content, use corresponding template structure.

After analysis, save report to: {NOTES_DIR}/{topic}/{Paper Title}.md

Note:
- Only write the paper note file itself, do NOT update _index.md or index files (main process handles this)
- In the "Knowledge Network" section, search workspace for related notes and add [[]] links
- Check reading list, mark entry as complete if found

Agent configuration: model: sonnet, subagent_type: general-purpose

B3: Collect Results & Unified Index Update

After all Agents complete:

1. Unified index update (avoid concurrent conflicts):
   • Read current _index.md, append all new paper entries at once
   • Read current index file, append all new entries at once
2. Reverse links: For all new papers, append reverse [[]] links in existing related paper notes
3. Run post-processing if post-archive.sh exists

B4: Cross-Paper Insights Summary

Generate a brief cross-paper insight (output directly to user, no separate file):

• Relationships between papers (complementary/competing/progressive)
• Common trends they point to
• 1-2 papers most worth diving deeper into

B5: Report Completion

Report to user: each paper's save path, assigned topic, number of links established.

---

Triage Mode Execution Flow

Enters this mode when input contains triage: prefix. Quick relevance assessment, no full notes generated.

TR1: Parse Input

Extract paper list and optional context: research direction description.

If no context provided and RESEARCH_FOCUS is configured, use it as the default research direction. Otherwise, ask the user to specify their research direction.

TR2: Quick Abstract Fetch

Fetch Abstract for each paper in parallel (only need abs page, not full text):

• arxiv ID → WebFetch https://arxiv.org/abs/{id}
• Other formats → WebSearch for abstract

TR3: Assess & Rank

For each paper output:

Paper	Relevance	Novelty	Practicality	Recommendation	Reason
{title}	High/Med/Low	High/Med/Low	High/Med/Low	Must-read / Worth-reading / Skip	{one-sentence reason}

Assessment dimensions:

• Relevance: Direct connection to the context research direction
• Novelty: New method/perspective, or incremental improvement
• Practicality: Directly useful inspiration for current work

TR4: Output Recommendations

Output results table sorted by recommendation, with suggestions:

• "Must-read" papers → suggest using /paper-xray batch: for batch processing
• "Skip" papers → explain skip reason

No files generated, no index updates. Pure conversation output.

---

Topic Research Mode Execution Flow

Enters this mode when input contains topic: or is judged to be a research topic.

T1: Parse Research Question

Extract: core topic, scope constraints (time/venue), paper count (default 5-8, n: to specify).

Break topic into 2-3 search dimensions.

T2: Multi-Round Search — Build Candidate Pool

Use WebSearch with multiple search dimensions in parallel (via Agent tool):

For each search dimension, launch an Agent to:
1. WebSearch with dimension-specific keywords
2. Collect arxiv IDs, titles, and brief descriptions
3. Return structured results

Merge all results, deduplicate, sort by relevance.

T3: Filter — Determine Analysis Paper List

Filtering principles:

• Coverage: Cover different sub-directions of the topic
• Representativeness: Prioritize foundational work + latest advances
• Diversity: At least one each of method/system/evaluation/survey
• Deduplication: Keep only the more representative of highly similar papers

Present filtered results to user (paper list + selection rationale), confirm before continuing.

T4: Parallel Deep Analysis

Launch independent Agent per paper, executing Single mode steps 1 through 6.7 (analysis only, no cognitive reflection).

Agent prompts include topic context, requiring analysis of paper's relationship to the topic.

Important: Each Agent only writes the paper note file, does NOT update index files (main process handles this).

T5: Synthesis — Generate Survey Report

After collecting all Agent results:

1. Technical approach comparison: Build comparison matrix
2. Field evolution: Timeline, inheritance relationships, technical route analysis, trends
3. Consensus & disagreements: Common conclusions, contradictions, blank directions
4. Research frontier assessment: Most promising directions, biggest unsolved problems, suggested entry points

T6: Archive

1. Unified index update: Batch update all _index.md and index files
2. Bidirectional links: Establish [[]] links between papers
3. Generate survey report: Save using references/survey-template.md to {NOTES_DIR}/{topic}/!survey-{topic-short}-{date}.md (! prefix ensures it sorts first in directory)
4. Run post-processing if post-archive.sh exists

---

Output Quality Standards

• Main body is objective analysis (steps 2-6), comprising 80%+ of the report
• Precise terminology, key concepts defined
• Mathematical formulas preserved and explained
• Experimental evaluation is critical, not just restating the results section
• Personal reflection is optional and never forced
• Structured presentation: tables, bullet points, comparisons — no essays