deepresearch

A multi-agent technical research workflow for Claude Code. Automatically researches GitHub repositories or technical articles through iterative AI agent coordination, and saves structured reports to Feishu (Lark) knowledge bases.

---

LLM Setup Guide

This section is written for AI assistants (Claude Code, etc.) to follow step-by-step when helping a user install this skill. Read the steps carefully, execute them in order, and verify each one before continuing.

Overview

Installing this skill requires four things:

#	Item	What it is
1	Skill files	Agent definitions and skill entrypoint, copied to ~/.claude/
2	lark-cli	CLI tool for Feishu API operations
3	DeepWiki MCP	MCP server for querying GitHub repo documentation
4	Config file	~/.deepresearch/lark-wiki.json with Feishu space details

---

Step 1 — Install skill files

Clone this repo and copy the files to your global Claude Code config directory:

git clone https://github.com/PPsteven/deepresearch /tmp/deepresearch

# agents
mkdir -p ~/.claude/agents
cp /tmp/deepresearch/agents/research-ask.md   ~/.claude/agents/
cp /tmp/deepresearch/agents/research-saver.md ~/.claude/agents/
cp /tmp/deepresearch/agents/researcher.md     ~/.claude/agents/

# skill entrypoint
mkdir -p ~/.claude/skills/research-lead
cp /tmp/deepresearch/skills/research-lead/SKILL.md ~/.claude/skills/research-lead/

Verify:

ls ~/.claude/agents/research-ask.md ~/.claude/agents/research-saver.md \
   ~/.claude/agents/researcher.md ~/.claude/skills/research-lead/SKILL.md

All four files must exist. If any are missing, re-run the copy commands above.

---

Step 2 — Install lark-cli

lark-cli is the command-line tool for Feishu API operations (reading/writing docs, comments, wiki nodes).

Install from the official repo: https://github.com/larksuite/cli

After installation, verify:

lark-cli --version

Then authenticate with Feishu:

lark-cli auth login

Follow the prompts to complete OAuth. When done, run:

lark-cli contact users me

If a user record is returned, authentication is working.

---

Step 3 — Install DeepWiki MCP

DeepWiki MCP lets Claude query GitHub repository documentation. It is required for the technical research workflow (Mode A).

Install the skill in Claude Code:

https://skills.sh/aahl/skills/mcp-deepwiki

After installation, verify by asking Claude Code:

"Use DeepWiki to look up the structure of github.com/anthropics/anthropic-sdk-python"

If it returns a list of documentation topics, the MCP is working.

---

Step 4 — Create config file

The skill reads Feishu space configuration from ~/.deepresearch/lark-wiki.json.

mkdir -p ~/.deepresearch
cp /tmp/deepresearch/lark-wiki.json ~/.deepresearch/lark-wiki.json

Open ~/.deepresearch/lark-wiki.json and fill in your Feishu knowledge base details:

{
  "spaces": {
    "My Space": {
      "space_id": "<your_space_id>",
      "space_url": "https://my.feishu.cn/wiki/<space_token>",
      "overview_doc": "<overview_doc_token>",
      "sections": {
        "code":    { "parent_node": "<node_token>", "label": "代码研究" },
        "article": { "parent_node": "<node_token>", "label": "技术文章" },
        "product": { "parent_node": "<node_token>", "label": "产品分析" }
      }
    }
  }
}

To find these values, open your Feishu knowledge base in a browser:

• space_id — visible in the wiki URL or via lark-cli wiki spaces list
• parent_node for each section — the node token of the folder you want to save into

---

Step 5 — Verify everything

Run this checklist before declaring the installation complete:

# 1. Skill files present
ls ~/.claude/agents/research-ask.md \
   ~/.claude/agents/research-saver.md \
   ~/.claude/agents/researcher.md \
   ~/.claude/skills/research-lead/SKILL.md

# 2. lark-cli installed and authenticated
lark-cli --version
lark-cli contact users me

# 3. Config file present and not empty placeholders
cat ~/.deepresearch/lark-wiki.json

If any check fails, go back to the corresponding step and resolve it before continuing.

---

Troubleshooting

Problem	Fix
lark-cli: command not found	Install lark-cli — see Step 2
lark-cli auth fails	Check network / Feishu account permissions
DeepWiki returns no results	Reinstall the MCP skill, restart Claude Code
~/.deepresearch/lark-wiki.json not found	Run Step 4
Config still has <placeholder> values	Edit ~/.deepresearch/lark-wiki.json with real values

---

Features

• Iterative Research — Up to 5 rounds of AI-driven questioning to ensure thorough coverage
• Multi-Agent Coordination — Specialized agents for research orchestration, execution, and document writing
• Parallel Research — Multiple researcher agents work simultaneously on independent questions
• Structured Outline — Research follows a fixed 8-section outline (overview, architecture, workflow, etc.)
• Feishu Integration — Automatically saves reports as multi-page wiki documents with Mermaid diagrams
• Document Optimization — Answers unresolved comments in Feishu docs and updates content accordingly

Architecture

User Input (GitHub URL / Article)
         │
         ▼
   research-lead (skill)
   ┌──────────────────────────┐
   │  Orchestrates workflow   │
   └──────────┬───────────────┘
              │ spawns
              ▼
   research-ask (agent)            ←── coordinates
   ┌──────────────────────────┐
   │  Iterative questioning   │
   │  Up to 5 rounds          │
   └──────────┬───────────────┘
              │ spawns (parallel)
              ▼
   researcher (agent)
   ┌──────────────────────────┐
   │  DeepWiki MCP queries    │
   │  WebFetch for articles   │
   └──────────────────────────┘
              │ reports back
              ▼
   research-saver (agent)
   ┌──────────────────────────┐
   │  Generates Mermaid docs  │
   │  Splits into sub-pages   │
   │  Writes to Feishu wiki   │
   └──────────────────────────┘

Agents

Agent	Role	Triggered by
research-lead	Workflow orchestrator, manages team lifecycle	User (via skill)
research-ask	Research coordinator, iterative questioning	research-lead
researcher	Research executor, DeepWiki + WebFetch queries	research-ask
research-saver	Document writer, saves to Feishu wiki	research-lead

Skills

Skill	Description
research-lead	Research workflows + Feishu doc optimization (two modes)

Usage

Research a GitHub repository

In Claude Code, trigger the research-lead skill:

研究 https://github.com/owner/repo 写入飞书

Or with explicit section targeting:

研究 https://github.com/owner/repo，写入"深度研究"知识库的 article 分类

Optimize Feishu document comments

优化文档 https://my.feishu.cn/wiki/<node_token>

This collects all unresolved comments across the document and its sub-pages, researches answers, updates the document content, and marks comments as resolved.

Research Output Structure

Research reports follow a fixed 8-section outline:

1. Project Overview — One-sentence positioning, core value proposition
2. Problem & Positioning — Pain points solved, why it's needed
3. Core Design Philosophy — Key design decisions, architectural choices
4. System Architecture — Module composition, data flow (with Mermaid diagrams)
5. Core Workflows — How main features work, critical paths
6. Tech Stack & Integration — Technology choices, installation
7. API / Command Reference — Complete interface or command listing
8. Comparison & Innovation — Side-by-side comparison with similar projects

Document Saving

research-saver converts reports into structured Feishu wiki pages:

• Single page for short reports (< 800 words, ≤ 2 sections)
• Multi-page with numbered sub-pages for standard reports
• Nested pages when any sub-page exceeds 1200 words
• Mermaid diagrams auto-generated for architecture and workflow sections
• Index entry appended to the space overview document

License

MIT