deep-paper-reading-skill

AI Sherlock / AI 福尔摩斯：只接手核心论文，专门拆证据链。

deep-paper-reading-skill is a public demonstration Skill for rigorous deep reading of one selected academic paper. It is designed to work after a first-pass triage step such as paper-triage-skill.

The idea is simple:

AI Watson -> selects the paper worth reading
AI Sherlock -> asks whether the claims are actually supported

Current status

This repository is a public Skill prototype and output-contract demo. It shows the workflow, example inputs, example outputs, and a handoff format for group-meeting preparation.

It is not a deterministic automation engine that automatically verifies any paper end-to-end. Treat it as an Agent-assisted deep-reading workflow: useful for structure, critique, and verification reminders, but not a replacement for reading the original paper.

What it does

• Reads one selected core paper or structured paper notes.
• Separates author claims, evidence, and reader inference.
• Builds a claim-evidence matrix.
• Flags weak support, possible overclaims, and places requiring original-paper verification.
• Reviews methods, data, figures, tables, limitations, and failure cases.
• Generates advisor / group-meeting questions.
• Creates a handoff file for group-meeting PPT preparation.

What it does not do

• It does not replace full-text reading.
• It does not fabricate data, citations, paper claims, figures, or conclusions.
• It does not guarantee that a paper is correct.
• It does not guarantee supervisor approval, grades, acceptance, or publication.
• It does not expose hidden scoring weights, private prompt chains, or internal quality gates.

Quick start

Copy or reference the Skill folder:

skills/deep-paper-reading-skill

Example request:

Use deep-paper-reading-skill.

Current task:
Prepare a 15-minute group meeting about one core paper.

Research direction:
AI-assisted literature review workflows for graduate students.

Paper material:
Paste the paper title, abstract, key method section, results summary, and important figure/table captions here.

Need:
Create a claim-evidence matrix, likely advisor questions, and a verification checklist.

Expected output:

• deep_reading_report.md
• claim_evidence_matrix.md
• figure_table_review.md
• advisor_questions.md
• verification_checklist.md
• handoff_to_group_meeting.md
• handoff_to_group_meeting.json

Example output contract

examples/sample-output/
  deep_reading_report.md
  claim_evidence_matrix.md
  figure_table_review.md
  advisor_questions.md
  verification_checklist.md
  handoff_to_group_meeting.md
  handoff_to_group_meeting.json

See:

• sample-core-paper-input.md
• deep_reading_report.md
• claim_evidence_matrix.md
• handoff_to_group_meeting.md

How it pairs with paper-triage-skill

candidate papers
  -> paper-triage-skill
  -> reading queue
  -> selected core paper
  -> deep-paper-reading-skill
  -> evidence audit + advisor questions
  -> group-meeting-ppt-skill

The public workflow is intentionally transparent but not overexposed. It shows the output shape and reading logic, but not hidden scoring weights or private prompt chains.

Tests

python -m unittest discover -s tests -v

License

MIT License. See LICENSE.