paper-reader

name	paper-reader
description	Read abstracts and, when available, full text to extract structured evidence for literature screening. Use when a paper is ambiguous after abstract screening or the user needs method, dataset, supervision, or limitation details.

paper-reader

Name

paper-reader

Description

Read abstracts and, when available, full text to extract evidence relevant to literature screening. This skill is for second-stage review, not just summarization.

Use When

• the user needs evidence from full text
• the workflow needs method, dataset, or supervision details
• a paper is ambiguous after abstract screening
• local PDFs or text files are available for deeper review

Inputs

• normalized paper metadata
• local file paths when present
• abstract text
• screening criteria

Outputs

• structured evidence entries
• abstract summary
• extracted method, dataset, task, supervision, and limitation notes
• explicit notice when full text is unavailable

Workflow

1. Start from existing metadata and abstract.
2. If full text is available locally, parse it.
3. Extract evidence for:
   • task
   • method
   • dataset
   • supervision or annotation
   • limitations
4. Attach evidence with section, excerpt, note, and confidence.
5. If only abstract is available, say so explicitly and keep confidence conservative.

Guardrails

• Do not pretend full text was read if it was not available.
• Distinguish abstract evidence from full-text evidence.
• Prefer short, attributable excerpts over vague claims.
• If parser output is poor or missing, mark the paper as needing manual review.

Online And Offline Behavior

• online: may retrieve metadata remotely, but evidence extraction should still depend on actual accessible text
• offline: use local PDFs, text extracts, or metadata files
• when full text is unavailable in either mode, report the gap explicitly

Example

Expected extracted fields:

{
  "method": "Transformer with gaze-supervised attention head",
  "dataset": "BDD-100K-derived driving videos",
  "supervision": "human gaze maps",
  "limitations": "No night-scene evaluation reported"
}