paper-notes_skill

Overview

This skill extracts structured, evidence-ready notes for each paper in a core set and writes them to papers/paper_notes.jsonl. It produces short, checkable fields (summary_bullets, method, key_results, limitations, bibkey, priority, evidence_level) rather than long prose. The output is optimized for downstream claim/citation and writing steps that require verifiable, citation-ready artifacts.

How this skill works

It iterates over papers/core_set.csv (optionally using outline/mapping.tsv and fulltext files) and creates one JSONL record per paper. High-priority papers are enriched with 3–6 summary bullets, explicit method description, concrete key_results (task+metric+setting), and specific limitations. The tool marks evidence_level as abstract or fulltext and emits an accompanying evidence_bank.jsonl of addressable snippets for citation.

When to use it

• You have papers/core_set.csv and need evidence-ready notes for survey writing.
• Before drafting survey sections that require verifiable claims and citations.
• When mapping highlights key papers for deeper extraction (priority high).
• After running dedupe-rank and finalizing the core set.
• When you need a searchable, machine-readable evidence bank for synthesis.

Best practices

• Enrich priority=high papers with fulltext when available (set evidence_level: "fulltext").
• Record results as task + metric + constraint (budget, data split, tool access) — numbers alone are insufficient.
• Write limitations as paper-specific caveats (protocol mismatch, missing ablations, unrealistic settings).
• Avoid copy-pasting generic limitations across records; vary wording and be concrete.
• Use outline/mapping.tsv to prioritize which papers to enrich first for maximum downstream value.

Example use cases

• Generate structured notes for a 100-paper core set, then run a claims step that cites specific evidence snippets.
• Produce an evidence_bank.jsonl with ≥7 items/paper on average to support automated figure/table generation.
• Refine priority=high classics (e.g., ReAct-style papers) with fulltext snippets for definitive method/results capture.
• Run in abstract-only mode for long-tail papers and fulltext for a curated subset to balance effort.