Get research paper

Get Research Paper

A research-discovery skill. Where the research-paper skill writes papers, this skill finds them. Give it a topic, get a ranked, de-duplicated reading list of real existing papers with verified DOIs, key findings, and ready-to-cite metadata.

This file is the entry point. Heavier guidance (per-source strategies, ranking criteria, summarization prompts) lives in topic folders and is loaded on demand.

---

1. When to activate

Slash commands

Command	What it does
/get-research-paper <topic>	Curated reading list (default 10 papers)
/find-paper <topic>	Alias for /get-research-paper
/find-papers <topic>	Alias for /get-research-paper
/fetch-paper <topic>	Alias for /get-research-paper
/papers-on <topic>	Alias for /get-research-paper
/scholar <topic>	Quick scholarly summary (5 papers, 2-line summaries)

Common options:

• --n <N> — number of papers (default 10).
• --years <range> — e.g. 2020-2024, last-5, since-2018.
• --source <src> — arxiv, scholar, pubmed, semantic-scholar, all (default).
• --depth <quick|standard|deep> — summary detail.
• --style <harvard|apa|ieee|...> — pre-format the bibliography.
• --audience <academic|technical|general> — adjust summary register.
• --handoff — emit a bibliography.yaml ready for the research-paper skill.

Natural-language patterns

• "get research paper on / about / for [topic]"
• "find research papers on [topic]"
• "find papers on / about [topic]"
• "what are the top papers on [topic]"
• "show me research on [topic]"
• "fetch papers about [topic]"
• "list papers on [topic]"
• "literature on [topic]" (shorter than /literature-review)
• "scholar [topic]"

Negative activation

Do NOT activate for:

• Requests to write a paper (route to research-paper).
• Requests to review or critique a draft (route to research-paper).
• Casual questions ("what is X?") that don't need scholarly sources.
• Pure code / API documentation lookup.

---

2. Output contract

Every run produces, at minimum:

1. Reading list — N papers with:
   • Title (full)
   • Authors (first 3 + "et al." if more)
   • Year
   • Venue / journal / preprint server
   • DOI / arXiv ID / URL
   • 2–4 sentence summary (problem → method → finding → significance)
   • Relevance score (1–5) and quality score (per citation_engine rubric)
   • Cite key (lowercase author_year_word) ready for use
2. Field briefing (optional, default ON for --depth deep) — a 1-paragraph synthesis of where the field is and what the dominant approaches are.
3. bibliography.yaml — canonical-format file ready to drop into the research-paper skill.
4. Known-gaps.md block — every paper that couldn't be verified is surfaced with severity and recommended fix.

See templates/reading-list.md, templates/paper-summary.md, templates/briefing.md.

---

3. Core principles

0. Anchor to TODAY's date FIRST. Before any search, determine today's actual date (via date -u +%Y-%m-%d, runtime context, or asking the user). Never default to training-cutoff dates. Year ranges like --years last-3 are computed from today. Full protocol: instructions/freshness.md.
1. Real papers only. Never invent papers, DOIs, authors, or findings. Use only sources the model can verify (or honestly mark [UNVERIFIED — offline]).
2. De-duplicate aggressively. Same DOI / arXiv ID / first-author + year + title prefix → one entry.
3. Rank by relevance and quality. A bad paper that mentions the topic is less useful than a great paper that's two clicks adjacent.
4. Cite-ready by default. Every entry has cite_key + DOI + ready-to-use formatted citation.
5. Triangulation. For load-bearing claims, prefer ≥ 2 independent sources. Note when a finding rests on a single source.
6. Honest about limits. Without web tools, the model relies on training-data knowledge — flag every entry accordingly.
7. Hand off cleanly. Output is consumable by the research-paper skill via --handoff mode.

---

4. Top-level workflow

intake → search-strategy → fan-out search → rank+dedupe →
verify → summarize → assemble briefing → output (+ optional handoff)

Each step has a dedicated playbook. Read the file for the step you're on; persist the artifact; move on. Master pipeline: workflows/search.md.

---

5. Source coverage

Source	When to prefer	Tool
arXiv	CS, ML, AI, physics, math, quant-bio	toolchains/arxiv_search.py (works offline-only via API)
Google Scholar	Generic / cross-discipline broad surveys	WebSearch with site:scholar.google.com
Semantic Scholar	API-friendly, citation graph, summaries	WebFetch of api.semanticscholar.org
PubMed / PubMed Central	Biomedical, life sciences	WebFetch of eutils.ncbi.nlm.nih.gov
DBLP	CS authors / venues / publication lists	WebFetch of dblp.org
ACM DL	HCI, systems, security, networks	WebSearch with site:dl.acm.org
IEEE Xplore	Engineering, signal, hardware	WebSearch with site:ieeexplore.ieee.org
OpenReview	NeurIPS, ICLR, ICML reviews + papers	WebFetch of openreview.net
Crossref	DOI verification + metadata fill-in	WebFetch of api.crossref.org
Retraction Watch	Retraction screening	WebFetch of retractionwatch.com / database

Per-source strategy details: sources/.

---

6. Ranking and quality

Each candidate paper is scored on:

• Authority (0–4) — venue quality (peer-review rigor, impact).
• Methodological rigor (0–3) — replicability, sample size, sound stats.
• Recency / relevance (0–3) — fresh + topical, OR foundational + canonical.
• Total (0–10) — used to rank.

Default reading lists keep papers scoring ≥ 5. Higher floors raise the bar (--quality-floor 7).

Full rubric: prompts/ranking.md (extends the citation_engine/source-evaluation.md of the research-paper skill).

---

7. Handoff to research-paper

After producing a reading list:

/get-research-paper "graph neural networks for fraud detection" \
    --n 25 --handoff --style ieee --years 2020-2024

Produces:

gnn-fraud-detection/
├── reading-list.md            # human-readable curated list
├── bibliography.yaml          # ← canonical file for research-paper skill
├── briefing.md                # 1-paragraph synthesis
└── Known-gaps.md              # any unverifiable items

The user then runs the writer skill with the produced bibliography:

/research "graph neural networks for fraud detection" \
    --style ieee --bibliography ./gnn-fraud-detection/bibliography.yaml

The writer reads the curated bibliography directly — no re-search needed.

---

8. Failure handling

• No web search available → use model-known papers, mark every entry [UNVERIFIED — offline], lower the recommended --n to 5–8, and surface the limitation in the briefing.
• Search returns nothing → broaden the query (drop adjectives, try synonyms), then return what was found with an honest note.
• Conflicting metadata across sources → prefer the published (peer-reviewed) version over the preprint; note the relationship.
• Retracted paper detected → drop from the list; flag in Known-gaps.md.
• Out-of-scope topic → surface a note in the briefing; deliver best-effort results.

---

9. Where to look next

• Plan a search → workflows/search.md
• Per-source strategy → sources/
• Ranking rubric → prompts/ranking.md
• Summarization → prompts/summarization.md
• Output templates → templates/
• Hand off to writer → workflows/handoff-to-writer.md
• arXiv search tool → toolchains/arxiv_search.py

This skill is intentionally smaller than the writer skill. Its job is discovery and curation; the heavy lifting (writing, methodology, review) lives in research-paper.