claims-extractor_skill

Overview

This skill extracts key claims, contributions, and assumptions from a manuscript and writes them to output/CLAIMS.md with explicit, locatable source pointers. It produces a split list of empirical vs conceptual claims so downstream checks can be automated and auditable. The output is designed for peer review, evidence audits, and structured downstream verification.

How this skill works

The skill reads output/PAPER.md end-to-end (respecting optional DECISIONS.md constraints) and identifies primary contributions, asserted claims, and underlying assumptions. Each item is normalized to a single sentence, labeled as empirical or conceptual, scoped, and annotated with a precise source pointer (section + page/figure/table + short quote). The final CLAIMS.md separates empirical and conceptual sections and flags underspecified claims or missing evidence.

When to use it

• Preparing or responding to peer review where claims must be auditable
• Performing an evidence audit or reproducibility check of a manuscript
• Generating a machine-readable artifact for downstream verification pipelines
• Before drafting rebuttals or revision plans to target specific assertions
• When you need to convert a paper into checklistable assertions for reviewers or CI tools

Best practices

• Ensure output/PAPER.md is the full plain-text manuscript; convert PDFs/HTML first
• If DECISIONS.md exists, apply its scope and format constraints before extraction
• Rewrite vague claims to include metric/dataset/baseline or mark them as underspecified
• Always include a precise source pointer: section + figure/table/page + short quote
• Differentiate empirical (data/experiment-backed) from conceptual (argument/definition-backed) claims

Example use cases

• Produce an auditable CLAIMS.md before submitting reviews to journals or conferences
• Feed empirical claims to automated experiment-checking skills to validate results
• Create a concise claims summary for authors preparing a revision plan
• Support reproducibility teams by listing assumptions that must hold for results to generalize
• Audit clinical or policy manuscripts where traceable assertions are required