deep-research - SKILL.md Agent Skill

name: deep-research description: Conducts iterative deep research on any topic using web search, progressive exploration, and structured synthesis. Use when asked for comprehensive research, deep investigation, thorough analysis, or multi-source exploration of any topic. Triggers: research, investigate, deep dive, comprehensive analysis, explore thoroughly, find everything about. allowed-tools: WebSearch, WebFetch, Write, Read, Glob, Bash

Deep Research Skill

Overview

This skill conducts systematic, multi-iteration research on any topic. Inspired by OpenAI's Deep Research and open-source implementations, it uses progressive exploration to build comprehensive understanding.

When to use: User asks for comprehensive research, investigation, deep analysis, or thorough exploration of a topic requiring multiple sources.

Research Methodology

Phase 1: Scope Definition

Before starting research, clarify:

Research Question: What exactly needs to be answered?
Key Aspects: What facets of the topic should be explored?
Parameters: Set depth and breadth based on complexity
- Simple topic: depth=2, breadth=3
- Moderate topic: depth=3, breadth=4 (default)
- Complex topic: depth=4-5, breadth=5-6

Create Research Directory:

./research/[topic-slug]/
├── findings.md      # Accumulated findings per iteration
├── sources.md       # All sources with URLs
└── report.md        # Final synthesized report

Phase 2: Iterative Research Loop

For each iteration (1 to depth):

Step 1: Generate Search Queries

Create targeted search queries based on:

Original research question
Current knowledge gaps (from previous iterations)
Unexplored angles discovered in findings
Contradictions requiring resolution

Query Generation Guidelines:

Make queries specific, not generic
Vary approach: definitional, comparative, recent news, expert sources
Build on previous learnings - don't repeat covered ground
Target entities, metrics, dates mentioned in findings
Include domain-specific terminology discovered

Generate breadth queries per iteration.

Step 2: Execute Searches

For each query:

Use WebSearch to find relevant sources
Review search results for relevance and quality
For the most promising 2-3 results per query, use WebFetch to get full content

Source Quality Signals:

Authoritative domains (.gov, .edu, major publications)
Recent publication dates for current topics
Expert authors or organizational sources
Primary sources over aggregators

Step 3: Extract Findings

From each source, extract:

Key Facts: Concrete, verifiable information
Data Points: Numbers, metrics, statistics with context
Entities: People, organizations, products, places mentioned
Dates/Timeline: When things happened or will happen
Quotes: Direct statements from experts or officials
Claims: Assertions that may need verification

Append to findings.md:

## Iteration [N] - [Date/Time]

### Query: "[search query]"

**Source: [Title]** ([URL])
- Finding 1
- Finding 2
- Key quote: "..."

**Source: [Title]** ([URL])
- Finding 1
...

Update sources.md:

[N]. [Title] - [URL] (accessed [date])

Step 4: Gap Analysis

After processing all queries in an iteration, assess:

Unanswered Questions: What aspects remain unaddressed?
Weak Areas: Single-source or low-confidence information?
Contradictions: Conflicting claims needing resolution?
Emerging Angles: New questions raised by findings?
Completeness Score: 1-10, how well is the question answered?

Decision Logic:

If completeness < 7 AND depth remaining → CONTINUE
If completeness >= 8 OR depth exhausted → PROCEED TO SYNTHESIS
If new searches aren't adding value (diminishing returns) → PROCEED TO SYNTHESIS

Step 5: Iterate or Proceed

If continuing:

Use gap analysis to inform next iteration's queries
Focus on weak areas and unanswered questions
Reduce breadth slightly for more targeted searches

If proceeding to synthesis:

Move to Phase 3

Phase 3: Synthesis

Step 1: Organize Findings

Read all accumulated findings from findings.md and organize by:

Logical themes/subtopics (NOT by iteration order)
Chronological order for historical topics
Categories relevant to the research question

Step 2: Reconcile Information

Identify consensus across multiple sources
Flag disputed or contradictory claims
Note confidence levels based on source quality and agreement
Distinguish facts from opinions/speculation

Step 3: Write Report

Create report.md with this structure:

# Research Report: [Topic]

**Generated**: [Date]
**Research Depth**: [N] iterations
**Sources Consulted**: [Count]

---

## Executive Summary

[2-3 paragraphs summarizing key findings, main conclusions, and critical insights]

---

## Key Findings

### [Subtopic 1]

[Synthesized findings with inline citations like [1], [2]]

### [Subtopic 2]

[Synthesized findings]

[Additional sections as needed]

---

## Analysis

[Synthesis of patterns, implications, and insights that emerge from the findings]

---

## Limitations & Gaps

- [What couldn't be determined]
- [Conflicting information that wasn't resolved]
- [Areas requiring further research]

---

## Sources

1. [Title] - [URL]
2. [Title] - [URL]
...

Parameters

Parameter	Default	Range	Description
depth	3	1-5	Number of research iterations
breadth	4	2-8	Queries per iteration
focus	none	string	Optional focus areas to prioritize

Usage Examples

Basic Research

"Use deep research to investigate the current state of quantum computing"

Configured Research

"Research autonomous vehicle regulations with depth=4 and breadth=5, focusing on US federal policy"

Comparative Research

"Deep research comparing React, Vue, and Svelte for enterprise applications"

Implementation Notes

Context Management

To avoid context overflow on deep research:

Save findings to files after each iteration
Read from files when synthesizing
Summarize previous iterations rather than carrying full context
Use the files as the source of truth

Progress Updates

After each iteration, provide a brief update:

Iteration 2/3 complete:
- Executed 4 queries
- Found 12 new findings
- Key discovery: [brief insight]
- Gaps remaining: [brief list]
- Next focus: [what iteration 3 will target]

Quality Checks

Throughout research:

Verify claims across multiple sources when possible
Flag single-source claims with lower confidence
Note recency of sources (important for current events)
Prefer primary sources over secondary coverage
Watch for circular sourcing (multiple articles citing same original)

Source Tracking

Maintain rigorous citation:

Every factual claim needs a source number
Record full URL for each source
Note access date for web sources
Include source title for readability

Error Handling

If WebSearch returns no results: try alternative query formulations
If WebFetch fails on a URL: note in findings, continue with other sources
If a topic has very limited sources: note this as a limitation
If sources heavily conflict: present both perspectives, don't force resolution

Output Structure

./research/[topic-slug]/
├── findings.md          # Raw findings from each iteration
│                        # Append-only during research
├── sources.md           # Numbered list of all sources
│                        # Updated after each search
└── report.md            # Final synthesized report
                         # Created in Phase 3

All files are markdown for easy reading and future reference.