Run Two-Phase Deep Research with Five Slash Commands

Ask an agent to "research X" and you get one long answer with no structure to extend or audit. Deep Research Skills, a 2k-star MIT project by Weizhena inspired by the RhinoInsight paper, splits research into two phases — an outline you can extend, then a deep investigation of each item — with human control at every stage.

Why This Skill Matters

The README positions the workflow for four kinds of work: academic research (paper surveys, benchmark reviews, literature analysis), technical research (technology comparison, framework evaluation, tool selection), market research (competitor analysis, industry trends, product comparison) and due diligence (company research, investment analysis, risk assessment). The same five skills serve all four: /research generates an outline of research items and the fields to collect for each, /research-add-items and /research-add-fields extend it, /research-deep researches every item with parallel agents, and /research-report generates a markdown report from the JSON results.

The repo ships those five skills four ways — research-en, research-zh, research-codex-en and research-codex-zh — so you pick a language and a runtime at install time.

Installation

git clone https://github.com/Weizhena/deep-research-skills.git
cd deep-research-skills

Claude Code takes the skills, a web-search agent with its modules that the README marks as required, and one Python dependency:

# English version
cp -r skills/research-en/* ~/.claude/skills/

# Chinese version
cp -r skills/research-zh/* ~/.claude/skills/

# Required: Install agent and modules
cp agents/web-search-agent.md ~/.claude/agents/
cp -r agents/web-search-modules ~/.claude/agents/

# Required: Install Python dependency
pip install pyyaml

OpenCode uses the same skills copy, but the README is emphatic that any model's web search there needs OPENCODE_ENABLE_EXA=1 — a plain export lasts one shell, so write it to ~/.bashrc; without it you only get web fetch, which the README calls weaker for the deep research phase. Its agent goes to ~/.config/opencode/agents/web-search.md, with the modules alongside. Codex gets skills/research-codex-en/* (or -zh) under ~/.codex/skills/, agents-codex/web-researcher.toml and the modules under ~/.codex/agents/, and a ~/.codex/config.toml with multi_agent = true and default_mode_request_user_input = true under [features] — bash scripts/install-codex.sh writes it for you, or you edit it by hand.

Real Workflow: A Landscape of AI Agent Demos

The README's own example topic:

1. Generate the outline:

/research AI Agent Demo 2025

You get a list of 17 AI agents to research — ChatGPT Agent, Claude Computer Use, Cursor and others — plus what information to collect for each. On Claude Code 2.1.0+ the /skill-name trigger works directly; older versions use run /skill-name. In Codex, pick the skill from /skills → List Skills, or ask naturally:

Use the research skill to build an outline for AI Agent Demo 2025

2. Not satisfied? Extend the outline before the deep phase:

/research-add-items
/research-add-fields

3. Run the deep phase:

/research-deep

The agent searches the web for every item and collects the detailed fields — the README's example lists company, release date, pricing, tech specs and reviews.

4. Generate the report:

/research-report

All the data becomes report.md, a complete markdown report with a table of contents, ready to read or share.

Tips

• /research-add-items and /research-add-fields both operate on the existing outline — they are the README's optional step between outline and deep research.
• The agent and modules are not optional. On every runtime the README labels an agent file, the web-search-modules directory and pip install pyyaml as required — web-search-agent.md for Claude Code, web-search-opencode.md for OpenCode, web-researcher.toml for Codex.
• Language is chosen at install time: research-en or research-zh for Claude Code and OpenCode, research-codex-en or research-codex-zh for Codex.
• The README's OpenCode section notes a default of gpt-5.4 and says any model's web search requires OPENCODE_ENABLE_EXA=1.
• Stuck? The README suggests asking your agent to explain the project: Help me understand this project: https://github.com/Weizhena/deep-research-skills.

When Not to Use This

The design is human-in-the-loop by intent — the README says it ensures precise control at every stage — and the README's walkthrough is separate commands you issue in turn — outline, optional extension, deep research, report. If you want one prompt that returns a finished report, this is not shaped that way. It is also a clone-and-copy install: the README documents no package or plugin path, and it marks the agent files and pyyaml as required rather than optional. On OpenCode without OPENCODE_ENABLE_EXA=1, the investigation step is down to web fetch.

See the leaderboard for more skills.