Stats
Actions
Tags
From commonplace
Deep paper breakdown (does not write vault notes). Use when user says 'analyze this paper', 'break down this paper', 'is this paper any good', or sends an arXiv ID/PDF with no save intent. Do NOT use when user wants the paper saved — that's wiki-ingest.
How this skill is triggered — by the user, by Claude, or both
Slash command
/commonplace:paper-analyzerThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Analyze research papers with intelligent extraction that never loads entire PDFs into context. Uses section-based extraction adapted to paper length, multi-agent analysis for complex papers, and quality scoring to ensure thorough coverage.
Analyze research papers with intelligent extraction that never loads entire PDFs into context. Uses section-based extraction adapted to paper length, multi-agent analysis for complex papers, and quality scoring to ensure thorough coverage.
Smart extraction adapts to paper structure. Even 100+ page papers stay within token budget because we extract only the critical sections (methods, results, experiments), never the full document.
When you receive a paper to analyze (arXiv ID, URL, or PDF path):
Acquire the paper if needed:
commonplace paper:fetch <arxiv-id-or-url>
Get paper info (minimal tokens):
commonplace paper:extract <pdf> info
Smart extraction (adapts to paper length):
commonplace paper:smart-extract <pdf>
Enrich metadata from external sources:
commonplace paper:enrich --arxiv-id <id>
Extract figure/table captions for reference:
commonplace paper:figures <pdf>
Analyze citations to understand the paper's context:
commonplace paper:citations <pdf>
Use single-agent analysis:
references/analysis_structure.md for the analysis frameworkassets/output_template.mdUse multi-agent analysis for deeper coverage:
paper-methodology-analyst and paper-results-interpreter agents in parallelpaper-reflection-agent to review quality (scores 0-100)Critical — agent context isolation: Subagents have completely isolated context windows. They cannot see the parent conversation, files you've read, or scripts you've run. Every agent prompt must be self-contained and include:
/tmp/papers/foo.pdf)commonplace commands directly — it's on PATH)Agents do have full tool access (Bash, Read, etc.) — the problem is they won't know what to operate on unless you tell them explicitly in the prompt.
The reflection loop is what makes analyses thorough — the reflection agent catches missing details, unsupported claims, and gaps that a single pass would miss.
When comparing multiple papers:
references/comparison_framework.mdcommonplace paper:compare <analysis1.md> <analysis2.md>
After writing an analysis, verify its completeness:
commonplace paper:quality <analysis.md> --min-score 60
The quality scorer checks for:
Target score: 80+ for production analyses. Scores below 60 indicate significant gaps.
Analyses use the template at assets/output_template.md. Key sections:
The TypeScript PDF extraction (pdfjs-dist) handles most academic papers well but has known issues with:
If extracted text looks garbled or truncated, note this in the analysis and work with whatever is available. The metadata enrichment (arXiv abstract, Semantic Scholar) can fill gaps.
references/analysis_structure.md — Detailed guide on structuring paper analyses with examplesreferences/comparison_framework.md — Framework for multi-paper comparative analysisassets/output_template.md — Markdown template for analysis outputnpx claudepluginhub noopz/stray-bits-sanctuary --plugin commonplaceCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.