deep-scientific-research - SKILL.md Agent Skill

name: deep-scientific-research description: > Deep scientific research pipeline with subagent orchestration, parallel web search, and citation management. Use this skill whenever the user asks for a literature review, scientific research summary, evidence synthesis, systematic review, or any deep research task that requires searching multiple sources and producing a cited report. Also triggers on phrases like "research this topic", "find papers on", "what does the evidence say about", "literature review on", or "deep dive into".

Deep Scientific Research Skill

A multi-agent pipeline for conducting rigorous scientific research using the Claude Agent SDK. The system breaks a research question into focused subtopics, spawns parallel researcher subagents to search the web, then synthesizes everything into a unified report with a numbered citation list.

Architecture

User Question
     |
     v
+-------------------+
|   Orchestrator    |  Plans subtopics, coordinates agents
|   (Lead Agent)    |
+---------+---------+
          | spawns 3-6 in parallel
    +-----+-----+-----+
    v     v     v     v
+------+------+------+
|Rsrch1||Rsrch2||Rsrch3|  Each searches web, returns
|      ||      ||      |  structured findings + URLs
+--+---++--+---++--+---+
   |       |       |
   +-------+-------+
           v
   +---------------+
   |  Synthesizer  |  Merges, deduplicates, renumbers
   |   Subagent    |  citations, writes final report
   +-------+-------+
           v
     Final Report
   (Markdown + References)

Subagent Definitions

This skill relies on two subagents that should be available in the project:

researcher

Tools: WebSearch, WebFetch
Model: sonnet (fast, cost-effective for search tasks)
Purpose: Searches the web for a specific subtopic. Returns structured findings with source URLs.
Output format: Markdown with ## Subtopic, ### Findings, ### Sources sections.

synthesizer

Tools: None (pure text synthesis)
Model: sonnet
Purpose: Takes all researcher outputs and produces a unified, well-structured report with globally renumbered citations.
Output format: Full research report with Executive Summary, numbered sections, Open Questions, and References list.

Workflow

When the user asks a research question:

PLAN - Break the research question into 3-6 focused subtopics. Each subtopic should be a specific, searchable angle of the question.
DELEGATE - For every subtopic, invoke the "researcher" subagent with a clear, precise prompt. Include the subtopic title and 2-3 specific search queries the researcher should try. Spawn researchers in PARALLEL whenever possible to save time.
SYNTHESIZE - Once all researchers report back, combine findings into a single, coherent research report using the "synthesizer" subagent. Pass ALL researcher outputs to the synthesizer verbatim.
DELIVER - Present the synthesizer's final report to the user. Save the report to out/report_<timestamp>.md.

Important Rules

Always use the researcher subagent for gathering information. Never search yourself.
Always use the synthesizer subagent for the final report. Never write the report yourself.
When calling a subagent, pass all necessary context in the prompt string - subagents cannot see this conversation.
If a researcher returns thin results, you may spawn a follow-up researcher with refined queries.
Your final message to the user should be the synthesized report, presented verbatim without modification.

Customization

Domain-specific research

The researcher subagent can be tuned to prioritize domain-specific databases:

Biomedical: PubMed, ClinicalTrials.gov, bioRxiv
Physics/Math: arXiv, APS journals
Computer Science: Semantic Scholar, ACM DL, DBLP
Legal: case law databases, government registers
Economics: NBER, SSRN, Fed publications

Cost Expectations

A typical research query spawns 3-6 researcher subagents, each making 3-5 web searches. Expect roughly:

Simple topic: ~$0.50-1.00 (3 researchers, ~15 searches)
Complex topic: ~$2.00-5.00 (6 researchers + follow-ups, ~30 searches)