Deep Research Skills
A structured deep research workflow for Claude Code, OpenCode, and Codex with human-in-the-loop control at every stage.
Evidence shows: README emphasizes human-in-the-loop design and requires user confirmation (default_mode_request_user_input = true), but lacks least-privilege details, such as whether skill files request unnecessary permissions. Data flow transparency is insufficient; no explanation of what data is collected and where it goes. No mention of sensitive data handling. Dependency is only pyyaml, but no version pinning or security audit. External effects include modifying config files and installing agents, but no rollback mechanism. Source attribution only mentions the inspiration paper, no internal citations. Deductions: missing least privilege, sensitive data handling, rollback, and detailed data flow.
Evidence shows: README and test file are mostly consistent in installation steps and configuration, but no error handling or failure messages. Dependency pyyaml is not version-pinned, availability unverified. Deductions: missing failure messages and dependency version pinning.
Evidence shows: Clearly lists multiple use cases (academic, technical, market, due diligence) and supports multiple environments (Claude Code, OpenCode, Codex). Capability boundaries not clearly stated, but trigger commands are clear. Environment fit is good with platform-specific installation instructions. Deductions: insufficient capability boundary description.
Evidence shows: README is well-structured, installation notes are detailed, command naming is stable, examples and FAQ sections are present. Known limitations not explicitly listed. MIT license present. No version number or changelog. Maintenance responsibility not clear. Deductions: missing versioning, changelog, and clear maintenance responsibility.
Evidence shows: Output is a markdown report, high usability. Marginal value lies in automating research workflow, but cost-benefit not quantified. Deductions: no cost-benefit analysis.
Evidence shows: README cites a paper, but no internal citations or cross-validation. Facts and inferences not clearly separated. Deductions: lack of traceability and cross-validation.
- No least-privilege details; skills may request unnecessary permissions.
- No data flow or sensitive data handling explanation; potential privacy risks.
- Dependencies are not version-pinned; supply chain risk.
- No rollback mechanism; difficult to undo installation.
- No version number or changelog; maintenance and updates are opaque.
What does this agent do, and when should you use it?
Deep Research Skills is a structured research workflow skill for Claude Code, OpenCode, and Codex, inspired by the RhinoInsight paper. It supports a two-phase research process: extensible outline generation and deep investigation. The skill set includes multiple SKILL files (research, research-add-items, research-add-fields, research-deep, research-report) and web search agents/modules. The human-in-the-loop design allows users to precisely control the research scope by adding or adjusting items and fields after outline generation. The final output is a Markdown report with a table of contents. Installation involves copying skill folders to the platform-specific config directories and installing the Python dependency pyyaml.
This skill package provides command-driven skills for Claude Code, OpenCode, and Codex. Users trigger /research to generate a research outline with items and field definitions, /research-add-items and /research-add-fields to extend the outline, /research-deep to conduct deep research where the AI sequentially searches the web for each item, and /research-report to convert JSON results into a Markdown report with a table of contents. It relies on web search agents (e.g., web-search-agent or web-researcher) and requires shell, network, and filesystem access, plus pyyaml. For Codex, configuration includes enabling multi_agent and web_researcher agent in config.toml.
- Academic researchers conducting paper surveys, benchmark reviews, or literature analysis who need a structured outline and deep information gathering.
- Engineers evaluating technologies or frameworks when making technical decisions, requiring detailed comparisons.
- Market analysts researching competitors or industry trends, needing to synthesize information from multiple sources.
- Investment or due diligence professionals assessing companies or projects, requiring systematic information collection and organization.
What are this agent's strengths and limitations?
- Human-in-the-loop design gives users precise control at every stage, avoiding unfocused research.
- Multi-platform support (Claude Code, OpenCode, Codex) covers major coding assistants.
- Two-phase workflow (outline generation + deep research) is clear and the outline is extensible.
- Output is a polished Markdown report with a table of contents, ready for sharing.
- Installation is manual and requires copying files to each platform's config directories.
- OpenCode requires OPENCODE_ENABLE_EXA=1 for full web search; otherwise only weaker web fetch is available.
- Codex setup requires editing config.toml and enabling multi_agent and web_researcher agent, which may be complex for some users.
- Requires Python and pyyaml, an additional dependency.
How do you install or deploy this agent?
Clone the repository: git clone https://github.com/Weizhena/deep-research-skills.git. For Claude Code, copy skills/research-en or research-zh to ~/.claude/skills/, copy agents/web-search-agent.md and agents/web-search-modules to ~/.claude/agents/, and install pyyaml. For OpenCode, set environment variable OPENCODE_ENABLE_EXA=1 (optionally persist in ~/.bashrc), copy skills to ~/.claude/skills/ (or research-zh), copy agents/web-search-opencode.md and modules to ~/.config/opencode/agents/. For Codex, create directories and copy skills to ~/.codex/skills/, copy agents-codex/web-researcher.toml and web-search-modules to ~/.codex/agents/, then add the agent configuration to ~/.codex/config.toml or run scripts/install-codex.sh.
How do you use this agent?
After installation, in Claude Code 2.1.0+ you can trigger skills directly with /research, e.g., '/research AI Agent Demo 2025'. After outline generation, use /research-add-items and /research-add-fields to adjust. Then run /research-deep for deep research, and /research-report to generate the final report. For older Claude Code versions, use 'run /research' format. In Codex, you can trigger skills from /skills -> List Skills or ask naturally, e.g., 'Use the research skill to build an outline for AI Agent Demo 2025'.
How does this agent compare with similar options?
The project is inspired by RhinoInsight but does not explicitly name other alternatives.