Tooluniverse Literature Deep Research

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:

# [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

Phase 1: Target Disambiguation + Profile (Default ON)

CRITICAL: This phase prevents "missing target details" when literature is sparse or noisy.

1.1 Resolve Official Identifiers

Use these tools to establish canonical identity:

UniProt_search → Get UniProt accession for human protein
UniProt_get_entry_by_accession → Full entry with cross-references
UniProt_id_mapping → Map between ID types
ensembl_lookup_gene → Ensembl gene ID, biotype
MyGene_get_gene_annotation → NCBI Gene ID, aliases, summary

Output for report:

## Target Identity

| Identifier | Value | Source |
|------------|-------|--------|
| Official Symbol | ATP6V1A | HGNC |
| UniProt | P38606 | UniProt |
| Ensembl Gene | ENSG00000114573 | Ensembl |
| NCBI Gene ID | 523 | NCBI |
| ChEMBL Target | CHEMBL2364682 | ChEMBL |

**Full Name**: V-type proton ATPase catalytic subunit A
**Synonyms/Aliases**: ATP6A1, VPP2, Vma1, VA68

1.2 Identify Naming Collisions

CRITICAL: Many gene names have collisions. Examples:

• TRAG: T-cell regulatory gene vs bacterial TraG conjugation protein
• WDR7-7: Could match gene WDR7 vs lncRNA
• JAK: Janus kinase vs Just Another Kinase
• CAT: Catalase vs chloramphenicol acetyltransferase

Detection strategy:

1. Search PubMed for "[SYMBOL]"[Title] - review first 20 titles
2. If >20% off-topic, identify collision terms
3. Build negative filter: NOT [collision_term1] NOT [collision_term2]

Output for report:

### Known Naming Collisions

- Symbol "ATP6V1A" is unambiguous (no major collisions detected)
- Related but distinct: ATP6V0A1-4 (V0 subunits vs V1 subunits)
- Search filter applied: Include "vacuolar" OR "V-ATPase", exclude "V0 domain" when V1-specific

1.3 Protein Architecture & Domains

Use annotation tools (not literature):

InterPro_get_protein_domains → Domain architecture
UniProt_get_ptm_processing_by_accession → PTMs, active sites
proteins_api_get_protein → Additional protein features

Output for report:

### Protein Architecture

| Domain | Position | InterPro ID | Function |
|--------|----------|-------------|----------|
| V-ATPase A subunit, N-terminal | 1-90 | IPR022879 | ATP binding |
| V-ATPase A subunit, catalytic | 91-490 | IPR005725 | Catalysis |
| V-ATPase A subunit, C-terminal | 491-617 | IPR022878 | Complex assembly |

**Length**: 617 aa | **Isoforms**: 2 (canonical P38606-1, variant P38606-2 missing aa 1-45)
**Active sites**: Lys-168 (ATP binding), Glu-261 (catalytic)

*Sources: InterPro, UniProt*

1.4 Subcellular Location

HPA_get_subcellular_location → Human Protein Atlas localization
UniProt_get_subcellular_location_by_accession → UniProt annotation

Output for report:

### Subcellular Localization

| Location | Confidence | Source |
|----------|------------|--------|
| Lysosome membrane | High | HPA + UniProt |
| Endosome membrane | High | UniProt |
| Golgi apparatus | Medium | HPA |
| Plasma membrane (subset) | Low | Literature |

**Primary location**: Lysosomal/endosomal membranes (vacuolar ATPase complex)
*Sources: Human Protein Atlas, UniProt*

1.5 Baseline Expression

GTEx_get_median_gene_expression → Tissue expression (TPM)
HPA_get_rna_expression_by_source → HPA expression data

Output for report:

### Baseline Tissue Expression

| Tissue | Expression (TPM) | Specificity |
|--------|------------------|-------------|
| Kidney cortex | 145.3 | Elevated |
| Liver | 98.7 | Medium |
| Brain - Cerebellum | 87.2 | Medium |
| Lung | 76.4 | Medium |
| Ubiquitous baseline | ~50 | Broad |

**Tissue Specificity**: Low (τ = 0.28) - broadly expressed housekeeping gene
*Source: GTEx v8*

1.6 GO Terms & Pathway Placement

GO_get_annotations_for_gene → GO annotations
Reactome_map_uniprot_to_pathways → Reactome pathways
kegg_get_gene_info → KEGG pathways
OpenTargets_get_target_gene_ontology_by_ensemblID → Open Targets GO

Output for report:

### Functional Annotations (GO)

**Molecular Function**:
- ATP hydrolysis activity (GO:0016887) [Evidence: IDA]
- Proton-transporting ATPase activity (GO:0046961) [Evidence: IDA]

**Biological Process**:
- Lysosomal acidification (GO:0007041) [Evidence: IMP]
- Autophagy (GO:0006914) [Evidence: IMP]
- Bone resorption (GO:0045453) [Evidence: IMP]

**Cellular Component**:
- Vacuolar proton-transporting V-type ATPase, V1 domain (GO:0000221) [Evidence: IDA]

### Pathway Involvement

| Pathway | Database | Significance |
|---------|----------|--------------|
| Lysosome | KEGG hsa04142 | Core component |
| Phagosome | KEGG hsa04145 | Acidification |
| Autophagy - animal | Reactome R-HSA-9612973 | mTORC1 regulation |

*Sources: GO Consortium, Reactome, KEGG*

Phase 2: Literature Search (Internal Methodology)

NOTE: This methodology is kept internal. The report shows findings, not process.

2.1 Query Strategy: Collision-Aware Synonym Plan

Step 1: High-Precision Seed Queries (Build Mechanistic Core)

Query 1: "[GENE_SYMBOL]"[Title] AND (mechanism OR function OR structure)
Query 2: "[FULL_PROTEIN_NAME]"[Title] 
Query 3: "[UNIPROT_ID]" (catches supplementary materials)

Purpose: Get 15-30 high-confidence, mechanistic papers that are definitely on-target.

Step 2: Citation Network Expansion (Especially for Sparse Targets)

Once you have 5-15 core PMIDs:

PubMed_get_cited_by → Papers citing each seed
PubMed_get_related → Computationally related papers  
EuropePMC_get_citations → Alternative citation source
EuropePMC_get_references → Backward citations from seeds

Citation-network first option: For older targets with deprecated terminology, citation expansion often outperforms keyword searching.

Step 3: Collision-Filtered Broader Queries

Broader query: "[GENE_SYMBOL]" AND ([pathway1] OR [pathway2] OR [function])
Apply collision filter: NOT [collision_term1] NOT [collision_term2]

Example for bacterial TraG collision:

"TRAG" AND (T-cell OR immune OR cancer) NOT plasmid NOT conjugation NOT bacterial

2.2 Database Tools

Literature Search (use all relevant):

• PubMed_search_articles - Primary biomedical
• PMC_search_papers - Full-text
• EuropePMC_search_articles - European coverage
• openalex_literature_search - Broad academic
• Crossref_search_works - DOI registry
• SemanticScholar_search_papers - AI-ranked
• BioRxiv_search_preprints / MedRxiv_search_preprints - Preprints

Citation Tools (with failure handling):

• PubMed_get_cited_by - Primary (NCBI elink can be flaky)
• EuropePMC_get_citations - Fallback when PubMed fails
• PubMed_get_related - Related articles
• EuropePMC_get_references - Reference lists

Annotation Tools (not literature, but fill gaps):

• UniProt_* tools - Protein data
• InterPro_get_protein_domains - Domains
• GTEx_* tools - Expression
• HPA_* tools - Human Protein Atlas
• OpenTargets_* tools - Target-disease associations
• GO_get_annotations_for_gene - GO terms

2.3 Full-Text Verification Strategy

WHEN TO USE: Abstracts lack critical experimental details (exact drugs, cell lines, concentrations, specific protocols).

Three-Tier Strategy:

Tier 1: Auto-Snippet Mode (Europe PMC) - FASTEST

Use for: Exploratory queries with 3-5 specific terms

results = EuropePMC_search_articles(
    query="bacterial antibiotic resistance evolution",
    limit=10,
    extract_terms_from_fulltext=["ciprofloxacin", "meropenem", "A. baumannii", "MIC"]
)

# Check which articles have full-text snippets
for article in results:
    if "fulltext_snippets" in article:
        # Snippets automatically extracted from OA full text
        for snippet in article["fulltext_snippets"]:
            # Use snippet["term"] and snippet["snippet"] for verification
            pass

Advantages:

• ✅ Single tool call (search + snippets)
• ✅ Bounded latency (max 3 OA articles, ~3-5 seconds total)
• ✅ No manual URL extraction
• ✅ Max 5 search terms

Limitations:

• ❌ Only works for OA articles with fullTextXML
• ❌ Limited to first 3 OA articles
• ❌ Europe PMC coverage only (~30-40% OA)

When to use: Initial exploration, quick verification of 1-2 papers

Tier 2: Manual Two-Step (Semantic Scholar, ArXiv) - TARGETED

Use for: Specific high-value papers you identified from search

# Step 1: Search
papers = SemanticScholar_search_papers(
    query="machine learning interpretability",
    limit=10
)

# Step 2: Extract from specific OA papers
for paper in papers:
    if paper.get("open_access_pdf_url"):
        snippets = SemanticScholar_get_pdf_snippets(
            open_access_pdf_url=paper["open_access_pdf_url"],
            terms=["SHAP", "gradient attribution", "layer-wise relevance"],
            window_chars=300
        )
        if snippets["status"] == "success":
            # Process snippets["snippets"]
            pass

ArXiv variant (100% OA, no paywall):

# All arXiv papers are freely available
snippets = ArXiv_get_pdf_snippets(
    arxiv_id="2301.12345",
    terms=["attention mechanism", "self-attention", "layer normalization"],
    max_snippets_per_term=5
)

Advantages:

• ✅ Full control over which papers to process
• ✅ Adjustable window size (20-2000 chars)
• ✅ Works for Semantic Scholar (~15-20% OA PDFs) and ArXiv (100%)
• ✅ Can process any number of papers

Limitations:

• ❌ Two tool calls per article (search → extract)
• ❌ Manual loop needed
• ❌ Slower than auto-snippet mode

When to use: Thorough review of key papers, preprint analysis

Tier 3: Manual Download + Parse (Fallback) - SLOWEST

Use for: Paywalled content via institutional access

# For paywalled PDFs accessible via institution
webpage_text = get_webpage_text_from_url(
    url="https://doi.org/10.1016/...",
    # Requires institutional proxy or VPN
)

# Extract relevant sections manually
if "Methods" in webpage_text:
    # Parse methods section
    pass

Limitations:

• ❌ Requires institutional access
• ❌ No snippet extraction (full HTML)
• ❌ Quality varies by publisher
• ❌ Slowest approach

When to use: Last resort for critical paywalled papers

Decision Matrix

Best Practices

1. Limit search terms to 3-5 specific keywords:

• ✅ Good: ["ciprofloxacin 5 μg/mL", "HEK293 cells", "RNA-seq"]
• ❌ Bad: ["drug", "method", "significant"] (too broad)

2. Check OA status before extraction:

if article.get("open_access") and article.get("fulltext_xml_url"):
    # Proceed with extraction
    pass

3. Adjust window size for context:

• Methods: 400-500 chars (full sentences)
• Quick verification: 150-200 chars
• Default: 220 chars (balanced)

4. Handle failures gracefully:

if "fulltext_snippets" not in article:
    # Fallback: use abstract or skip
    print(f"No full text available: {article['title']}")

5. Document full-text sources in report:

## Methods Verification

**Antibiotic concentrations** (verified from full text):
- Study A: Ciprofloxacin 5 μg/mL [PMC12345, Methods section]
- Study B: Meropenem 8 μg/mL [arXiv:2301.12345, Experimental Design]

*Note: Full-text verification performed on 8/15 OA papers (53% coverage)*

2.5 Tool Failure Handling

Automatic retry strategy:

Attempt 1: Call tool
If timeout/error:
  Wait 2 seconds
  Attempt 2: Retry
If still fails:
  Wait 5 seconds  
  Attempt 3: Try fallback tool
If fallback fails:
  Document "Data unavailable" in report

Fallback chains:

2.6 Open Access Handling (Best-Effort)

If Unpaywall email provided: Check OA status for all papers with DOIs

If no Unpaywall email: Use best-effort OA signals:

• Europe PMC: isOpenAccess field
• PMC: All PMC papers are OA
• OpenAlex: is_oa field
• DOAJ: All DOAJ papers are OA

Label in report:

*OA Status: Best-effort (Unpaywall not configured)*

Phase 3: Evidence Grading

CRITICAL: Grade every claim by evidence strength to prevent low-signal mentions from diluting the report.

Evidence Tiers

How to Apply

In report, label sections and claims:

### Mechanism of Action

ATP6V1A is the catalytic subunit responsible for ATP hydrolysis in the V-ATPase 
complex [★★★ Mechanistic: PMID:12345678]. Loss-of-function mutations cause 
vacuolar pH dysregulation [★★★: PMID:23456789].

The target has been implicated in mTORC1 signaling through lysosomal amino acid 
sensing [★★☆ Functional: PMID:34567890], though direct interaction data is limited.

A genome-wide screen identified ATP6V1A as essential in cancer cell lines 
[★☆☆ Association: PMID:45678901, DepMap].

Theme-Level Grading

For each theme section, summarize evidence quality:

### 3.1 Lysosomal Acidification (12 papers)
**Evidence Quality**: Strong (8 mechanistic, 3 functional, 1 association)

[Theme content...]

Report Structure: Mandatory Completeness Checklist

CRITICAL: This checklist/template applies to Full Deep-Research Mode. For Factoid / Verification Mode, use a short fact-check report (see Phase 0) and do not force the full 15-section template.

Output Files

1. [topic]_report.md - Main narrative report (Full Deep-Research Mode)
2. [topic]_factcheck_report.md - Short verification report (Factoid / Verification Mode)
3. [topic]_bibliography.json - Full deduplicated bibliography (always created)
4. methods_appendix.md - Methodology details (ONLY if user requests)

Report Template

# [TARGET/TOPIC]: Comprehensive Research Report

*Generated: [Date]*
*Evidence cutoff: [Date]*
*Total unique papers: [N]*

---

## Executive Summary

[2-3 paragraphs synthesizing key findings across all sections]

**Bottom Line**: [One-sentence actionable conclusion]

---

## 1. Target Identity & Aliases
*[MANDATORY - even for non-target topics, clarify scope]*

### 1.1 Official Identifiers
[Table of IDs or scope definition]

### 1.2 Synonyms and Aliases  
[List all known names - critical for complete literature coverage]

### 1.3 Known Naming Collisions
[Document collisions and how they were handled]

---

## 2. Protein Architecture
*[MANDATORY for protein targets; state "N/A - not a protein target" otherwise]*

### 2.1 Domain Structure
[Table of domains with positions, InterPro IDs]

### 2.2 Isoforms
[List isoforms, functional differences if known]

### 2.3 Key Structural Features
[Active sites, binding sites, PTMs]

### 2.4 Available Structures
[PDB entries, AlphaFold availability]

---

## 3. Complexes & Interaction Partners
*[MANDATORY]*

### 3.1 Known Complexes
[List complexes the protein participates in]

### 3.2 Direct Interactors
[Table of top interactors with evidence type and scores]

### 3.3 Functional Interaction Network
[Describe network context]

---

## 4. Subcellular Localization
*[MANDATORY]*

[Table of locations with confidence levels and sources]

---

## 5. Expression Profile
*[MANDATORY]*

### 5.1 Tissue Expression
[Table of top tissues with TPM values]

### 5.2 Cell-Type Expression
[If single-cell data available]

### 5.3 Disease-Specific Expression
[Expression changes in disease contexts]

---

## 6. Core Mechanisms
*[MANDATORY - this is the heart of the report]*

### 6.1 Molecular Function
[What the protein does biochemically]
**Evidence Quality**: [Strong/Moderate/Limited]

### 6.2 Biological Role
[Role in cellular/organismal context]
**Evidence Quality**: [Strong/Moderate/Limited]

### 6.3 Key Pathways
[Pathway involvement with evidence grades]

### 6.4 Regulation
[How the target is regulated]

---

## 7. Model Organism Evidence
*[MANDATORY]*

### 7.1 Mouse Models
[Knockout/knockin phenotypes, if any]

### 7.2 Other Model Organisms
[Yeast, fly, zebrafish, worm data if relevant]

### 7.3 Cross-Species Conservation
[Conservation and functional studies]

---

## 8. Human Genetics & Variants
*[MANDATORY]*

### 8.1 Constraint Scores
[pLI, LOEUF, missense Z - with interpretation]

### 8.2 Disease-Associated Variants
[ClinVar pathogenic variants]

### 8.3 Population Variants
[gnomAD notable variants]

### 8.4 GWAS Associations
[Any GWAS hits for the locus]

---

## 9. Disease Links
*[MANDATORY - include evidence strength]*

### 9.1 Strong Evidence (Genetic + Functional)
[Diseases with causal evidence]

### 9.2 Moderate Evidence (Association + Mechanism)
[Diseases with supporting evidence]

### 9.3 Weak Evidence (Association Only)
[Diseases with correlation/association only]

### 9.4 Evidence Summary Table

| Disease | Evidence Type | Score | Key Papers | Grade |
|---------|---------------|-------|------------|-------|
| [Disease 1] | Genetic + Functional | 0.85 | PMID:xxx | ★★★ |
| [Disease 2] | GWAS + Expression | 0.45 | PMID:yyy | ★★☆ |

---

## 10. Pathogen Involvement
*[MANDATORY - state "None identified" if not applicable]*

### 10.1 Viral Interactions
[Any viral exploitation or targeting]

### 10.2 Bacterial Interactions
[Any bacterial relevance]

### 10.3 Host Defense Role
[Role in immune response if any]

---

## 11. Key Assays & Readouts
*[MANDATORY]*

### 11.1 Biochemical Assays
[Available assays for target activity]

### 11.2 Cellular Readouts
[Cell-based assays and phenotypes]

### 11.3 In Vivo Models
[Animal models and endpoints]

---

## 12. Research Themes
*[MANDATORY - structured theme extraction]*

### 12.1 [Theme 1 Name] (N papers)
**Evidence Quality**: [Strong/Moderate/Limited]
**Representative Papers**: [≥3 papers or state "insufficient"]

[Theme description with evidence-graded citations]

### 12.2 [Theme 2 Name] (N papers)
[Same structure]

[Continue for all themes - require ≥3 representative papers per theme, or state "limited evidence"]

---

## 13. Open Questions & Research Gaps
*[MANDATORY]*

### 13.1 Mechanistic Unknowns
[What we don't understand about the target]

### 13.2 Therapeutic Unknowns
[What we don't know for drug development]

### 13.3 Suggested Priority Questions
[Ranked list of important unanswered questions]

---

## 14. Biological Model & Testable Hypotheses
*[MANDATORY - synthesis section]*

### 14.1 Integrated Biological Model
[3-5 paragraph synthesis integrating all evidence into coherent model]

### 14.2 Testable Hypotheses

| # | Hypothesis | Perturbation | Readout | Expected Result | Priority |
|---|------------|--------------|---------|-----------------|----------|
| 1 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | HIGH |
| 2 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | HIGH |
| 3 | [Hypothesis] | [Experiment] | [Measure] | [Prediction] | MEDIUM |

### 14.3 Suggested Experiments
[Brief description of key experiments to test hypotheses]

---

## 15. Conclusions & Recommendations
*[MANDATORY]*

### 15.1 Key Takeaways
[Bullet points of most important findings]

### 15.2 Confidence Assessment
[Overall confidence in the findings: High/Medium/Low with justification]

### 15.3 Recommended Next Steps
[Prioritized action items]

---

## References

*[Summary reference list in report - full bibliography in separate file]*

### Key Papers (Must-Read)
1. [Citation with PMID] - [Why important] [Grade: ★★★]
2. ...

### By Theme
[Organized reference lists]

---

## Data Limitations

- [Any databases that failed or returned no data]
- [Any known gaps in coverage]
- [OA status method used]

*Full methodology available in methods_appendix.md upon request.*

Extended Reference: For detailed tool tables, examples, and templates, read REFERENCE.md in this skill directory. The agent can access it via: read skills/tooluniverse-literature-deep-research/REFERENCE.md