菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | paper-analyzer |
| description | 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. |
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> --save /tmp/<paper-slug>-extracted.txt
--save writes the extracted section text to a plain text file — reused later by the groundedness check in the reflection step (see paper-reflection-agent)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:
--save (gets nearly everything)references/analysis_structure.md for the analysis frameworkassets/output_template.mdUse multi-agent analysis for deeper coverage:
--savepaper-methodology-analyst and paper-results-interpreter agents in parallelpaper-reflection-agent to review quality (scores 0-100) — pass it the --save file path so it can run the groundedness checkCritical — 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 outputAnswer questions from vault notes. Use when user asks 'how does X relate to Y', 'what do the papers say about Z', 'compare approaches to W', or discusses concept connections. Also fire immediately after wiki-ingest when user pivots from saving to asking how the new source relates to other vault topics. Also fire as a pre-ingest relevance check: before newly-shared content is dismissed as not vault-worthy, check whether it connects to existing vault notes. Also fire proactively as a conversation develops: when discussion drifts near vault topics, surface a genuinely non-obvious connection the vault holds — sparingly, only when it clears the bar. Do NOT answer from memory — read the actual notes.
ALWAYS use when the user wants a source saved to the vault — pastes an arXiv ID/URL, says 'add this to the vault', 'save this', 'remember this', 'I found this interesting', or describes a specific finding/technique/project to capture. Do NOT activate for general topic conversation — only when there's explicit save intent or a concrete source (URL, paper ID, pasted content).
Autonomous vault improvement loop (writes changes). Use when user says 'improve the vault', 'autoimprove', 'fix what you can', 'what's the vault score'. Do NOT use for read-only diagnostics — that's wiki-lint.
Fill concept stubs with citation-grounded definitions. Use when user says 'fill in the stubs', 'compile concepts', 'flesh out concepts', or as follow-up after wiki-lint reports stubs. Do NOT use for new sources not yet in vault — that's wiki-ingest.
User-invoked: surface hidden concept connections via local embeddings (Ollama + nomic-embed-text). Flags related-but-unlinked (sentence, concept) pairs, then adds wikilinks where the connection holds. Invoke explicitly via /wiki-deep-link.
Set up, list, rename, or merge research domains. Use when user says 'add a domain for X', 'set up a domain', 'what domains exist', 'change scope of X domain', 'merge X and Y domains'. Do NOT speculatively fire on ingest prompts — wiki-ingest hands off explicitly when no domain matches.