A structured, parallel-first research skill that enables AI agents to decompose complex queries, execute concurrent searches, and synthesize high-confidence answers.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install phy-research-deep
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install phy-research-deep using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Deep Research?
Deep Research is a highly efficient, parallel-first research methodology designed for AI agents. Instead of forcing agents into slow, sequential, and redundant search-read-search loops that take minutes to resolve, this skill enables a structured approach that completes complex research in under 60 seconds. By deploying Openclaw Skills like Deep Research, developers can upgrade their agent workflows to perform smart query decomposition, parallel search execution, selective fetching of high-value sources, and objective synthesis.
This methodology is highly tool-agnostic, easily integrating with various search MCPs or page fetch tools. It ensures that the AI agent approaches any query with analytical depth, actively pursuing opposing perspectives to eliminate confirmation bias, and rating its own confidence based on the strength and consensus of the evidence gathered.
Deep Research Use Cases
- Multi-Faceted Investigations: Resolving complex questions that require cross-referencing information from multiple disparate sources.
- Market & Competitive Analysis: Conducting deep dives into competitors, comparing technologies, and analyzing market trends.
- Fact-Checking & Source Verification: Investigating claims, establishing confidence levels, and detecting inconsistencies in public reports.
- Rapid Technology Appraisals: Answering "what is the current state of X" using fresh, time-bound web lookups.
- Automated Intelligence Gathering: Reducing the latency of gathering reliable data for decision-making agents.
How Deep Research Works
- Decompose (Query Breakdown): The agent divides a vague user prompt into 3-5 precise, non-overlapping search strings. This includes at least one adversarial query to combat confirmation bias.
- Parallel Search: Executes all sub-queries concurrently using a web search tool in a single round trip, avoiding slow sequential lookups.
- Selective Fetch: Evaluates snippets and fetches only the top 2-3 most authoritative, information-dense primary sources instead of scraping everything.
- Synthesize & Cite: Aggregates findings into a standardized markdown report featuring Key Findings, a Confidence Assessment matrix, identified Gaps & Unknowns, and fully cited Sources.
Deep Research Setup
This skill is tool-agnostic and integrates into your AI agent's prompt instructions or system profile. Simply append the core methodology instructions to your agent configuration.
To configure an environment utilizing standard Openclaw Skills search components, make sure you have your Model Context Protocol (MCP) servers set up for web searches and fetching:
# Configure your MCP settings to support parallel web search and web fetch tools
{
"mcpServers": {
"websearch-mcp": {
"command": "npx",
"args": ["-y", "@openclaw/mcp-websearch"]
}
}
}
Deep Research Data Schema & Taxonomy
The Deep Research skill utilizes a standardized markdown schema to structure research findings. This ensures consistent data organization across search results:
| Section | Description | Sub-Components |
|---|---|---|
| Key Findings | Synthesized evidence-backed assertions | Assertion, direct inline citations, source links |
| Confidence Assessment | Logical proof table with validation metrics | Finding name, Confidence Rating (High/Medium/Low), Basis |
| Gaps & Unknowns | Unverified information or missing data | Open questions, potential biases, blind spots |
| Sources | Full directory of referenced materials | Title, URL, one-line relevance note |
Confidence Rating Taxonomy:
- High: 3+ independent sources agree, includes primary data, no contradictions.
- Medium: 2 sources agree or 1 highly authoritative source, minor explainable contradictions.
- Low: Single source, conflicting evidence, or outdated material (>12 months).
Deep Research Advanced Features
- Adversarial Query Generation: Auto-injects negative-case lookups (e.g., "criticisms", "problems") to prevent confirmation bias.
- Tool-Agnostic Extensibility: Easily swaps standard search tools for academic engines (e.g., arXiv, Semantic Scholar) or developer forums (e.g., Reddit, Hacker News).
- Anti-Pattern Guardrails: Built-in guardrails block sequential lookup traps, cap max page fetches to 3 to optimize token usage, and prohibit summary-of-summaries reporting.
- Time-Bound Targeting: Explicitly appends current temporal markers (e.g., the current year) to prevent stale search data.
SKILL.md
Loading
Related Openclaw Skills
NotebookLM Research Assistant
An automated browser-driven skill that bridges Claude Code and Google NotebookLM to query your documents and retrieve citation-backed Gemini answers.
Accessibility (A11y) Reviewer
An AI-powered web accessibility reviewer that automatically audits web pages and code structures against WCAG 2.2 AA compliance standards.
UI Reviewer
An AI-powered frontend code reviewer that automates design system compliance, structural audits, responsiveness checks, and visual consistency for React and Vue applications.
Project Truth Audit
A comprehensive whole-repository audit skill for AI agents to contrast stated system claims with the underlying reality of files, tests, and runtime evidence.
Frontend Design
An expert frontend design skill that guides AI agents to construct high-quality, production-grade web interfaces and components that avoid generic AI aesthetics.
Content Compound
An automated content atom library builder that scans past content to extract reusable ideas, claims, and data points, eliminating blank-page creator burnout.
Featured*
