literature-deep-research

Files (3)

• SKILL.md
• run.cpython-313.pyc · scripts/__pycache__/
• run.py · scripts/

SKILL.md (first lines)

---
name: tooluniverse-literature-deep-research
description: Conduct comprehensive literature research with target disambiguation, evidence grading, and structured theme extraction. Creates a detailed report with mandatory completeness checklist, biological model synthesis, and testable hypotheses. For biological targets, resolves official IDs (Ensembl/UniProt), synonyms, naming collisions, and gathers expression/pathway context before literature search. Default deliverable is a report file; for single factoid questions, uses a fast verification mode and may include an inline answer. Use when users need thorough literature reviews, target profiles, or to verify specific claims from the literature.
---

# Literature Deep Research Strategy (Enhanced)

A systematic approach to comprehensive literature research that **starts with target disambiguation** to prevent missing details, uses **evidence grading** to separate signal from noise, and produces a **content-focused report** with mandatory completeness sections.

**KEY PRINCIPLES**:
1. **Target disambiguation FIRST** - Resolve IDs, synonyms, naming collisions before literature search
2. **Right-size the deliverable** - Use *Factoid / Verification Mode* for single, answerable questions; use full report mode for “deep research”
3. **Report-first output** - Default deliverable is a report file; an inline answer is allowed (and recommended) for Factoid / Verification Mode
4. **Evidence grading** - Grade every claim by evidence strength (mechanistic paper vs screen hit vs review vs text-mined)
5. **Mandatory completeness** - All checklist sections must exist, even if "unknown/limited evidence"
6. **Source attribution** - Every piece of information traceable to database/tool
7. **English-first queries** - Always use English terms for literature searches and tool calls, even if the user writes in another language. Only try original-language terms as a fallback if English returns no results. Respond in the user's language

---

## Workflow Overview

```
User Query
  ↓
Phase 0: CLARIFY + MODE SELECT (factoid vs deep report)
  ↓
Phase 1: TARGET DISAMBIGUATION + PROFILE (default ON for biological targets)
  ├─ Resolve official IDs (Ensembl, UniProt, HGNC)
  ├─ Gather synonyms/aliases + known naming collisions
  ├─ Get protein length, isoforms, domain architecture
  ├─ Get subcellular location, expression, GO terms, pathways
  └─ Output: Target Profile section + Collision-aware search plan
  ↓
Phase 2: LITERATURE SEARCH (internal methodology, not shown)
  ├─ High-precision seed queries (build mechanistic core)
  ├─ Citation network expansion from seeds
  ├─ Collision-filtered broader queries
  └─ Theme clustering + evidence grading
  ↓
Phase 3: REPORT SYNTHESIS
  ├─ Progressive writing to [topic]_report.md
  ├─ Mandatory completeness checklist validation
  └─ Biological model + testable hypotheses
  ↓
Optional: methods_appendix.md (only if user requests)
```

---

## Phase 0: Initial Clarification

### Mandatory Questions

1. **Target type**: Is this a biological target (gene/protein), a general topic, or a disease?
2. **Scope**: Is this a *single factoid to verify* (“Which antibiotic?”, “Which strain?”, “Which year?”) or a comprehensive/deep review?
3. **Known aliases**: Any specific gene symbols or protein names you use?
4. **Constraints**: Open access only? Include preprints? Specific organisms?
5. **Methods appendix**: Do you want methodology details in a separate file?

### Mode Selection (CRITICAL)

Pick exactly one mode based on the user’s intent and the question structure:

1. **Factoid / Verification Mode** (single concrete question; answer should be a short phrase/sentence)
2. **Mini-review Mode** (narrow topic; 1–3 pages of synthesis)
3. **Full Deep-Research Mode** (use the full template + completeness checklist)

**Heuristic**:
- If the user asks “X has been evolved to be resistant to *which antibiotic*?” → **Factoid / Verification Mode**
- If the user asks “What does the literature say about X?” → **Full Deep-Research Mode**

### Factoid / Verification Mode (Fast Path)

**Goal**: Provide a correct, source-verified single answer, with minimal but explicit evidence attribution.

**Deliverables** (still file-backed):
1. `[topic]_factcheck_report.md` (≤ 1 page)
2. `[topic]_bibliography.json` (+ CSV) containing the key paper(s)

**Fact-check report template**:
```markdown
# [TOPIC]: Fact-check Report

*Generated: [Date]*
*Evidence cutoff: [Date]*

## Question
[User question]

## Answer
**[One-sentence answer]** [Evidence: ★★★/★★☆/★☆☆/☆☆☆]

## Source(s)
- [Primary paper citation: journal/year/PMID/DOI as available]

## Verification Notes
- [1–3 bullets: where in the paper the statement appears (Abstract/Results/Methods), and any key constraints]

## Limitations
- [If full text not available, or if only review evidence exists]
```

**Required verification behavior**:
- Prefer ToolUniverse literature tools (Europe PMC / PubMed / PMC / Semantic Scholar) over general web browsing.
- Use full-text snippet verification when possible (Europe PMC auto-snippet tier is ideal).
- Avoid adding extra claims (e.g., “not X”) unless the paper explicitly supports them.

**Suggested tool pattern**:
- `EuropePMC_search_articles(query=..., extract_terms_from_fulltext=[...])` to pull OA full-text snippets for the key terms.
- If OA snippets unavailable: fall back to `PMC_search_papers` (if in PMC) or `SemanticScholar_search_papers` → `SemanticScholar_get_pdf_snippets`.

**Evidence grading (factoid)**:
- If the statement is explicitly made in a primary experimental paper (Results/Methods/Abstract): label **T1 (★★★)**.
- If it’s only in a review: label **T4 (☆☆☆)** and try to locate the primary source.

### Detect Target Type

| Query Pattern | Type | Action |

From lamm-mit/scienceclaw, Apache-2.0.

More from lamm-mit/scienceclaw