Deep Research Skill for Claude Code / OpenCode / Codex

A structured research workflow skill for Claude Code, OpenCode, and Codex, supporting two-phase research: outline generation (extensible) and deep investigation. Human-in-the-loop design ensures precise control at every stage.

Use Cases

• Academic Research: Paper surveys, benchmark reviews, literature analysis
• Technical Research: Technology comparison, framework evaluation, tool selection
• Market Research: Competitor analysis, industry trends, product comparison
• Due Diligence: Company research, investment analysis, risk assessment

Core Capabilities

• Two-Phase Research: Generate an outline first, then conduct deep investigation.
• Extensible Outline: Add more research items or fields to the existing outline.
• Parallel Deep Research: Automatically search the web for each item using parallel agents.
• Markdown Report: Generate a complete report with table of contents.
• Human-in-the-Loop: You control each step and can refine the outline before deep research.

How It Works

1. Phase 1: Generate Outline – Tell the skill your topic. It creates a structured research list with fields to collect.
2. Phase 2: Deep Research (optional) – Add more items or fields if needed, then run deep research. The skill searches the web for each item automatically.
3. Phase 3: Generate Report – All collected data is compiled into a single Markdown report ready to read or share.

Commands

Older versions of Claude Code use run /research format. Codex users can trigger skills from /skills -> List Skills or ask naturally.

Example

Topic: AI Agent Demo 2025

1. /research AI Agent Demo 2025 – Generates a list of 17 AI Agents to research (ChatGPT Agent, Claude Computer Use, Cursor, etc.) and what info to collect for each.
2. (Optional) /research-add-items or /research-add-fields – Add more research items or field definitions.
3. /research-deep – AI automatically searches the web for each item, one by one. You get detailed info for each Agent (company, release date, pricing, tech specs, reviews...).
4. /research-report – All data is compiled into report.md – a complete markdown report with table of contents.

Platform Requirements

Claude Code

• Works with Claude Code 2.1.0+ (direct command trigger) or older versions (use run /skill-name).
• Requires Python package pyyaml.

OpenCode

• Default model: gpt-5.4.
• Must set environment variable OPENCODE_ENABLE_EXA=1 for web search (otherwise only web fetch is available, which is weaker). It is recommended to make it persistent by adding to ~/.bashrc.
• Requires Python package pyyaml.

Codex

• Requires adding the following configuration to ~/.codex/config.toml:

suppress_unstable_features_warning = true

[features]
multi_agent = true
default_mode_request_user_input = true

[agents.web_researcher]
description = "Use this agent when you need to research information on the internet, particularly for debugging issues, finding solutions to technical problems, or gathering comprehensive information from multiple sources. This agent excels at finding relevant discussions. Use when you need creative search strategies, thorough investigation, or compilation of findings from multiple sources."
config_file = "agents/web-researcher.toml"

• The web-researcher.toml agent file and web-search-modules directory from the repository must be placed in the agent configuration path.
• Requires Python package pyyaml.

Getting Help

If you have questions, ask Claude Code, OpenCode, or Codex to explain this project:

Help me understand this project: https://github.com/Weizhena/deep-research-skills

References

• RhinoInsight: Improving Deep Research through Control Mechanisms for Model Behavior and Context

License

MIT