Literature Close Read

Source: https://github.com/aipoch/medical-research-skills

Literature Close Reading

When to Use

• When you have a full paper converted from PDF to Markdown and need a structured, in-depth interpretation rather than a brief abstract-style summary.
• When you must extract reproducible experimental details (datasets, settings, controls, metrics, statistics) for replication or reimplementation.
• When you need to map the paper's logical chain (motivation → problem → method → experiments → conclusions) and identify missing links or ambiguities.
• When you want a systematic list of limitations, threats to validity, and follow-up research questions grounded strictly in the text.
• When figures/tables are referenced via Markdown images and you need them incorporated into the interpretation without guessing beyond what is shown.

Key Features

• Reads the entire Markdown paper text, prioritizing Methods and Results for technical fidelity.
• Produces a structured close-reading report in Markdown (UTF-8), following a predefined template.
• Extracts and organizes:
  • research background and problem statement
  • methodological details and experimental design
  • key results and statistical evidence (as explicitly stated)
  • limitations and threats to validity
  • reproducible points and follow-up questions
• Supports Markdown inputs that include pagination headers like ## Page XX and image references such as ![page-01](...).
• Enforces a strict constraint: summarize only what is explicitly present in the text/images; do not infer or speculate.
• Uses external guidance and templates:
  • Requirements/checklist: references/guide.md
  • Output template: assets/deep_reading_template.md

Dependencies

• pdf-extract (version: not specified) — used only when the source is PDF and must be converted to Markdown first.

Example Usage

# 1) (Optional) Convert PDF to Markdown if you only have a PDF
# Note: exact command/options depend on your local pdf-extract installation.
pdf-extract paper.pdf > paper.md

# 2) Run the close-reading process (manual or via your orchestration tool):
# Input: paper.md (full text converted from PDF, may include `## Page XX` and images)
# Guidance: references/guide.md
# Template: assets/deep_reading_template.md

# 3) Save the final report as UTF-8 Markdown under outputs/
mkdir -p outputs
# Example output file name:
# outputs/paper_close_reading.md

Minimal expected I/O contract:

• Input: a single .md file containing the full paper text (PDF-to-Markdown), optionally with:
  • page headers like ## Page 01
  • image references like ![page-01](...)
• Output: one UTF-8 encoded .md report saved to outputs/, formatted according to assets/deep_reading_template.md.
• Language: default output is Chinese; if the user specifies a language, output in that language.

Implementation Details

• Input reading rules

  • Treat the Markdown as the authoritative source of truth.
  • Pagination markers (e.g., ## Page XX) may be used for navigation and citation, but should not alter meaning.
  • Image references may be used to interpret figures/tables only to the extent that the content is explicitly visible/legible.

• Extraction and summarization rules

  • Focus on Methods and Results first; then connect to background, problem statement, and conclusions.
  • Capture experimental details precisely: datasets, splits, baselines, ablations, hyperparameters, training/inference settings, evaluation metrics, and statistical tests—only if stated.
  • If a required field in the template cannot be filled from the text, write "Not specified".

• Quality constraints

  • No speculation: do not add assumptions, unstated motivations, or inferred mechanisms.
  • Maintain traceability: ensure each claim in the report can be traced back to explicit paper content (text or figure/table).
  • Output must be valid Markdown and saved in UTF-8 to avoid encoding issues.

• Files used

  • Requirements and checklist: references/guide.md
  • Output template: assets/deep_reading_template.md
  • Output directory: outputs/ (create if missing)

When Not to Use

• Do not use this skill when the required source data, identifiers, files, or credentials are missing.
• Do not use this skill when the user asks for fabricated results, unsupported claims, or out-of-scope conclusions.
• Do not use this skill when a simpler direct answer is more appropriate than the documented workflow.

Required Inputs

• A clearly specified task goal aligned with the documented scope.
• All required files, identifiers, parameters, or environment variables before execution.
• Any domain constraints, formatting requirements, and expected output destination if applicable.

Agent Execution Workflow

Follow these steps in order when the user provides a paper for close reading.

Step 1: Validate Input

• Confirm the user has provided the paper content (paste, file path, or PDF path).
• If PDF, inform the user it must be converted to Markdown first.
• Required: The paper text as Markdown. Optional: specific focus areas.

Step 2: Read and Parse the Full Text

• Read the entire Markdown content. Use ## Page XX markers for navigation.
• Identify major sections: Introduction, Methods, Results, Discussion, Limitations.
• Prioritize Methods and Results for detailed extraction.

Step 3: Extract Methods Details

• Capture: datasets, splits, baselines, ablations, hyperparameters, settings, metrics, tests.
• If any field is not explicitly stated, write "Not specified" — do NOT infer.
• Record exact values as stated.

Step 4: Extract Results and Evidence

• Capture key quantitative results (metrics, scores, p-values, confidence intervals).
• Note which figures/tables contain the supporting data.
• Report only what is explicitly stated or clearly visible.

Step 5: Identify Limitations

• Extract each stated limitation from the Limitations section.
• Note obvious unstated limitations (small sample, single-center, etc.).
• Distinguish author-stated from critically-identified.

Step 6: Fill the Report Template

• Use assets/deep_reading_template.md as your output structure.
• Fill each section with extracted information.
• For missing info, write "Not specified" — never fabricate.
• Default output language: Chinese. Override if user specifies another.

Step 7: Quality Check

• Verify every claim traces back to explicit paper content.
• Ensure no speculative content was added.
• Confirm valid Markdown, UTF-8 encoded.
• Save to outputs/literature_close_read_result.md.

Output Contract

• Return a structured deliverable that is directly usable without reformatting.
• If a file is produced, prefer a deterministic output name such as literature_close_read_result.md unless the skill documentation defines a better convention.
• Include a short validation summary describing what was checked, what assumptions were made, and any remaining limitations.

Validation and Safety Rules

• Validate required inputs before execution and stop early when mandatory fields or files are missing.
• Do not fabricate measurements, references, findings, or conclusions that are not supported by the provided source material.
• Emit a clear warning when credentials, privacy constraints, safety boundaries, or unsupported requests affect the result.
• Keep the output safe, reproducible, and within the documented scope at all times.

Failure Handling

• If validation fails, explain the exact missing field, file, or parameter and show the minimum fix required.
• If an external dependency or script fails, surface the command path, likely cause, and the next recovery step.
• If partial output is returned, label it clearly and identify which checks could not be completed.

Input Validation

This skill accepts requests that match the documented purpose of literature-close-read and include enough context to complete the workflow safely.

Do not continue the workflow when the request is out of scope, missing a critical input, or would require unsupported assumptions. Instead respond:

literature-close-read only handles its documented workflow. Please provide the missing required inputs or switch to a more suitable skill.

Quick Validation

Run this minimal verification path before full execution when possible:

No local script validation step is required for this skill.

Expected output format:

Result file: literature_close_read_result.md
Validation summary: PASS/FAIL with brief notes
Assumptions: explicit list if any

User Checkpoints

• Before executing batch processing, overwriting files, long-running searches, or multi-stage generation, confirm scope and output format with the user.
• Before proceeding when a key judgment is ambiguous, evidence is insufficient, or the workflow is entering the next stage, confirm with the user.