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

Reference Retrieval Skill

Core Function

This skill helps users quickly retrieve and filter high-quality, highly relevant references from PubMed based on natural language descriptions.

Workflow

Step 1: Construct Boolean Query

1. Analyze intent: Understand user semantics, extract core medical keywords (MeSH Terms preferred).
2. Construct Boolean query:
   • Logical operators: Use AND (intersection), OR (union), NOT (exclusion).
   • Truncation: Use * to match variants (e.g., diagno* matches diagnosis, diagnostic).
   • Example: For "metformin treatment of type 2 diabetes", construct: Metformin AND "Diabetes Mellitus, Type 2" AND (Therapy OR Treatment).

[!TIP] About Escaping: If the query contains double quotes " (e.g., phrase search), you must wrap the entire query in single quotes or escape internal double quotes with \" to avoid command-line parsing errors.

• Correct: python scripts/pubmed_search.py 'Metformin AND "Type 2 Diabetes"'
• Correct: python scripts/pubmed_search.py "Metformin AND \"Type 2 Diabetes\""

Step 2: Execute Search

Prefer LitSense for semantic search; only use PubMed Boolean search when precise control or complex Boolean logic is needed.

Scenario A: Semantic/Natural Language Search (LitSense) [Preferred] Suitable for direct natural language queries or finding literature related to complex statements.

python scripts/litsense_search.py "natural language query"

Scenario B: Basic Boolean Search (PubMed) Suitable when a precise Boolean query has been constructed.

python scripts/pubmed_search.py "YOUR_BOOLEAN_QUERY"

• --max: Returns 20 articles by default; specify --max 50 for more.

Step 3: Evaluation & Iteration

If initial results are inadequate (too few, low relevance, or empty), must automatically iterate, up to 5 rounds.

1. Analyze cause:
   • Zero results: Keywords may be misspelled or too specific.
   • Irrelevant results: Keywords are ambiguous or Boolean logic is incorrect.
2. Refine query:
   • Broaden scope: Remove non-essential AND conditions, add OR synonyms.
   • Narrow scope: Add qualifiers (e.g., diagnosis, therapy), or use field tags [ti].
   • Switch API: If LitSense search fails or lacks precision, try constructing a precise Boolean query for PubMed.
3. Iterate:
   • Repeat the cycle: modify query, run script, evaluate results.
   • Limit: Maximum 5 retries. If still no satisfactory results after 5 attempts, report tried strategies and request more information from the user.

Step 4: Filter & Present

After the script returns JSON results, filter 3-5 best references based on these criteria:

1. Selection criteria:

   • Relevance: Title/abstract must directly address the user's question.
   • Article type: Prefer Review, Systematic Review, Meta-Analysis. Then high-quality RCT or Original Article. Never cite Letters, Editorials, or Comments.
   • Recency: Prefer literature from the past 5-10 years (exceptions for classic foundational works).
   • Open Access (OA): If looking for OA/free literature, check for is_oa: true or pmcid field in results, indicating free full text via PMC.

2. Output format: Present results strictly in the following structure:

   I. Citation Marking Add citation markers at key points in the response content (or in user-provided text).

   • Format example: ...effectively improved prognosis [1]. or ...treatment efficacy [1].
   • Language should match the user's query language.

   II. References List references in citation order, format as follows:

   [1] PMID: 34479503 | Title. Journal. https://pubmed.ncbi.nlm.nih.gov/34479503/
   
   Citation rationale: Briefly explain why this reference supports the stated point.
   
   [2] ...

   PubMed link generation rules:

   • Always provide direct link: https://pubmed.ncbi.nlm.nih.gov/{pmid}
   • For example, PMID 34479503 becomes: https://pubmed.ncbi.nlm.nih.gov/34479503

   [!NOTE] Example: [1] The augment of regulatory T cells undermines the efficacy of anti-PD-L1 treatment in cervical cancer. BMC Immunol. https://pubmed.ncbi.nlm.nih.gov/34479503/

Important Notes

• Accuracy: A query that is too broad yields irrelevant results; too narrow may yield nothing. If results are few, try reducing search terms.
• Language consistency: Never auto-translate. Summary language must match the user's query language.
• Search language: Always translate non-English keywords to accurate English medical terms for searching (searches must use English), but present final results matching the user's language.

When to Use

• Use this skill when the user explicitly needs to perform the core task of reference-retrieval-skill and has provided the minimum executable input.
• Use this skill when you need a structured deliverable rather than general advice.
• Use this skill when the current task can be completed using this skill's bundled scripts, templates, or reference materials.

When Not to Use

• Do not proceed when required input files, identifiers, parameters, or context are missing — ask the user to provide them first.
• Do not assume capabilities beyond this skill's declared scope when the user requests external operations or inferences.
• Do not proceed without user confirmation when overwriting existing results, executing high-cost batch operations, or expanding task scope.

Required Inputs

Output Contract

• Primary output: Structured result or target file aligned with this skill's objective.
• Optional output: Intermediate check notes, issue list, supplementary suggestions, or generated file paths.
• Format requirement: Unless the user specifies otherwise, prefer stable, reviewable Markdown or JSON; if the skill's bundled script requires a fixed format, use that format.
• If partially complete: Must explicitly mark as PARTIAL and state which steps are completed and which remain.

Failure Handling

• Missing critical input: Explicitly state which fields, files, or identifiers are missing and pause.
• Script, template, or resource execution failure: Report the failing step, likely cause, and recovery suggestions — do not silently degrade.
• Partial completion only: Return the verified portion first, then list remaining blockers and suggested next steps.

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.

Input Validation

This skill accepts requests that match the documented purpose of reference-retrieval-skill 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:

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

Quick Validation

• Check that key scripts, templates, or reference file paths this skill depends on exist.
• Check that the final output contains the core fields, sections, or files specified for this task.
• Check that results clearly mark assumptions, limitations, and incomplete items.