Deep Research

Decompose complex research questions into sub-questions, research each with live web search, synthesize findings, verify conclusions, and output a structured research report.

---

中文说明

Overview

Deep Research is a Claude Code / Reasonix skill that transforms any complex research question into a structured, verifiable research report. It follows a rigorous 5-step methodology: decompose → research → synthesize → verify → output.

Unlike a simple "ask AI and get an answer" workflow, Deep Research:

• Decomposes the question into 3–7 independent sub-questions
• Researches each sub-question with real-time web search (up to 3 in parallel)
• Synthesizes findings into core discoveries and a preliminary conclusion
• Verifies the conclusion by deliberately searching for counter-evidence
• Iterates up to 2 rounds if verification fails
• Outputs a structured markdown report with full source citations

How It Works

User Question
     ↓
Step 1: Decompose → 3~7 sub-questions
     ↓
Step 2: Research each sub-question (parallel, max 3 per batch)
     ↓
Step 3: Synthesize → preliminary conclusion
     ↓
Step 4: Verify → design falsifiable checkpoints → search for counter-evidence
     ↓
Step 5: Pass? → Generate report file
        Fail? → Back to Step 1 (max 2 iterations)

Step 1 — Decompose

Break the question into 3–7 orthogonal sub-questions following this progression: definition/concept → current state/facts → causes/mechanisms → impact/trends → solutions/comparisons.

Step 2 — Research

For each sub-question, run web searches and optionally web_fetch key pages. Extract facts, data, and viewpoints. Every sub-question gets a structured answer with cited sources.

Search degradation strategy: If 2 consecutive searches return only irrelevant homepage results, fall back to direct web_fetch on authoritative domains (official docs > reputable blogs > community wikis).

Step 3 — Synthesize

Cross-reference all sub-question answers. Identify 3–5 core findings. Form a preliminary conclusion. Flag uncertainties and information gaps explicitly.

Step 4 — Verify

Design 3–5 falsifiable verification checkpoints — claims that can be proven wrong. Search for counter-evidence, opposing views, or more authoritative data. Each checkpoint gets ✅ Pass or ❌ Fail with reasoning.

Step 5 — Output

Scenario	Action
Verification passes	Generate structured report file (deep-research-<topic-slug>.md)
Verification fails, iterations < 2	Return to Step 1 with adjusted decomposition
Max iterations reached, still failing	Generate report with ⚠️ annotations on unresolved issues

Report Structure

Generated reports follow a standardized 6-section format with full source traceability:

Section	Content
1. Sub-Question Breakdown	Decomposition table
2. Per-Question Research	Search results, sources, and structured answers for each sub-question
3. Preliminary Synthesis	3–5 core findings, preliminary conclusion, information gaps
4. Verification	Falsifiable checkpoints, counter-evidence search, per-checkpoint pass/fail
5. Final Conclusion	Revised answer after verification (or annotated with unresolved issues)
6. Source Index	All cited sources with URLs

File Structure

deep-research/
├── SKILL.md                # Skill definition
├── report-template.md      # Report output template
├── README.md               # This file
└── README_CN.md            # Chinese readme

Installation

1. Copy the deep-research/ directory into your Reasonix (or Claude Code) skills folder:

   • Project-scoped: <project>/.reasonix/skills/deep-research/
   • Global: ~/.reasonix/skills/deep-research/

2. The skill will appear in your Skills index on the next session launch (or /new).

Usage

Invoke via the skill system:

/deep-research What are the latest advancements in solid-state batteries?

Or programmatically:

/run-skill deep-research "Compare Rust vs Zig for systems programming in 2025"

The skill spawns as a subagent, runs the full research pipeline independently, and returns the path to the generated report file.

Requirements

Requirement	Detail
Platform	Claude Code, Reasonix (or compatible agent runtime)
Tools	web_search, web_fetch, write_file, read_file
Model	Recommended: deepseek-v4-pro; minimum: deepseek-v4-flash

Design Principles

Principle	How It's Applied
Decompose before research	Prevents shallow answers to complex questions
Every claim needs a source	All factual assertions must cite URLs
Verify by seeking disconfirmation	Actively search for counter-evidence, not just confirmation
Iterate, don't fabricate	Up to 2 rounds; failures are annotated, never invented
Structured output	Consistent report format enables cross-topic comparison