Paper Search Pro
Multi-source literature search with adjustable depth. Four tiers, five data sources.
When to use this skill
- User wants to find academic papers
- User is preparing a literature review, systematic review, scoping review, or meta-analysis
- User wants to scope research on a topic for a thesis, proposal, coursework, or news story
- User asks what research exists on a topic
When NOT to use
- User wants to read a specific paper (use PDF reader)
- User wants to summarize a single known paper (use a summarizer)
- User wants to download PDFs given DOIs
- User already has a literature set and wants to write a review
Architecture
The main agent drives the workflow per this SKILL.md. Python helpers handle deterministic work with no LLM inside and no external API key required.
Data sources:
- L1 OpenAlex (primary) - deep top-100 multi-strategy
- L2 PubMed (medical) - MeSH enricher
- L2 arXiv (CS/preprint) - freshness sentinel
- L3 Semantic Scholar - influentialCitationCount + abstract fallback
- L3 CrossRef - funder / license / clinical-trial-number
Classification uses Inline SubAgents (parallel, file-IPC, 5 per message).
Output includes HTML report, Markdown, BibTeX, RIS, CSV, and PRISMA-S log.
The 4 tiers
- Quick (~5-8 min, 20-60 papers): fast scope
- Standard (~10-17 min, 60-180 papers): scope a topic, write background
- Deep (~30-45 min, 180-400 papers): thorough review writing
- Audit (~2-3 hr, 400-1000+ papers): systematic review, PRISMA, Cochrane, meta-analysis
Execution discipline
Four invariants govern every step:
- A: NEVER cd into the Skill directory. Run helpers from user PWD with PYTHONPATH.
- B: Dispatch classifier SubAgents in parallel (up to 5 in one message).
- C: Announce every skip with what, why, what is lost, and how to recover.
- D: Read a step cited reference when that step is non-trivial.
Workflow steps
- Setup - resolve install path and working directory
- Plan the query - detect language, PICO/SPIDER/PEO framework, journal-rank intent
- Retrieve from OpenAlex - multi-strategy deep crawl
- Run L2 boosters - PubMed enrich, arXiv freshness, NSSD, yiigle
- Federate - dedup and merge all sources
- Classify in parallel batches - LLM SubAgents with RCS rubric
- Compute saturation curve
- Decide next action - expand citations or stop
- Expand citations if applicable
- Enrich top-N papers with Semantic Scholar and CrossRef
- Write executive summary
- Render report - HTML, Markdown, exports
- Write PRISMA-S log
- Open report in browser and report to user
Output files
All outputs are relative to the user working directory:
- report.html - Main deliverable
- report.md - Markdown copy
- papers.csv, papers.bib, papers.ris, papers.json - Exports
- kg_classified.json - Internal KG with RCS scores
- summary.md - Executive summary
- execution_log.json - PRISMA-S log
- chart_data.json, paper_list.json, metadata.json - Sibling data
- curve.json - Saturation snapshot
- query_plan.json - Search plan
Error handling
- Config missing keys: Direct user to setup reference, halt
- Rate limit: Helper auto-retries; if persistent, drop that enricher
- OpenAlex 404 on DOI: Use title search fallback
- L2 booster returns 0 papers: Skip silently, note in PRISMA-S log
- SubAgent invalid JSON: 5-layer fallback parser
- HTML output size: No size cap, handles 10+ MB HTML cleanly
