Opencode Deep Research Report Generation Skill
Traditional deep research is expensive, industry reports cost $50–500+ each, making them unaffordable for individual researchers and indie developers. AI-driven research is often too shallow (a few summaries then a conclusion) or fabricates numbers. This project uses a 4-stage pipeline (outline → data collection → parallel writing → validation and assembly) to produce deep research reports matching the standards of brokerage firms and third-party research institutes, taking ~10–20 minutes per report at a cost of < ¥0.6.
https://github.com/hoolulu/deep-research…

hoolulu/deep-research

Source: https://github.com/hoolulu/deep-research

deep-research Skill

🇨🇳 中文 · 🇬🇧 English

Deep Research Report Generation Skill · Supports 19 languages output

Current version: See updates (https://github.com/hoolulu/deep-research/commits/main)

✨ One Minute Overview

The entire interaction and search for materials is conducted in the specified language—not merely a translation of the output.

I. Why You Need This

If you’ve tried getting AI to do research, you’ve probably run into these pitfalls:

Search + summary → too shallow, only a few snippets with no depth.
Industry reports priced per copy $50–500+ → too expensive for individuals.
Overseas tools → can’t search domestic resources such as Baidu Baike, Zhihu, 199IT, iResearch.
AI inventing numbers → looks plausible but has no source.

This skill completes a 4-layer process before delivering the report. It doesn’t just search and output; it follows: analyze → search & verify → write → verify.

II. Who Should Use It

Indie developers, independent researchers, small teams.
Anyone who needs professional-grade research capabilities but doesn’t want to rely on paid databases or research institutions.

III. Output of One Standard Mode Research

The above figures are typical ranges for standard mode; actual results may vary depending on topic complexity, data availability, search engine responsiveness, etc.

📂 View all generated reports → — Browse all generated research reports here; click to open and read.

IV. How It Works

The entire process consists of 4 stages, executed automatically in sequence:

1 Analyze outline — analyze the topic, generate research framework and search plan  
↓  
2 Collect data — SearXNG / Exa cascading search → Scrapling batch scraping → data pool extraction → data quality check  
↓  
3 Parallel writing — multiple chapters written simultaneously, facts embedded directly in prompts, no tool calls  
↓  
4 Validate and assemble — batch validate → assemble-report → convert-citations → qa-report  

V. Search Pipeline & Built-in Resources

All tools are built-in—no additional purchases needed. The system uses a three-layer cascading search strategy:
SearXNG (author-deployed meta search engine, 70+ engines including Baidu/Google/Brave) → Exa (OMO built-in cold backup) → over ten free search engines + domestic data sources (final fallback). Each layer checks if it’s reachable before proceeding; if available, it’s used immediately without probing lower layers, ensuring fast response. The entire flow:

Layer 1 — SearXNG (author-deployed, 70+ engines including Baidu/Google/Brave, ready to use out of the box)  
↓  
If unavailable: Layer 2 — Exa (OMO built-in cold backup, zero cost)  
↓  
If unavailable: Layer 3 — Free source reinforcement (fallback)  
├─ Search reinforcement line  
│   Known source reinforcement line  
├─ DuckDuckGo  
│   Baidu Baike / Wikipedia  
├─ Bing China  
│   Zhihu / 36Kr / The Paper  
├─ Brave / Mojeek  
│   199IT / iResearch / East Money  
├─ Semantic Scholar  
│   National Bureau of Statistics / Weibo / CSDN  
└─ GDELT / arXiv  
    Douban / Huxiu  

The search reinforcement line can dynamically discover any other websites, not limited to the list above. All source URLs are eventually batch-scraped for full text by Scrapling.

VI. Unique Report Highlights

VII. Three Depth Levels

Parameters are in profiles.json; modify and restart to apply. Character count excludes spaces and Markdown syntax.

VIII. Screenshots

(Placeholder for screenshots)

IX. Installation

🧠 Method 1: AI Easy Install (Recommended)

Copy the following prompt into the OpenCode chat and send it—AI will do everything automatically:

Please research the project at https://github.com/hoolulu/deep-research, and follow the documentation requirements in order:  
1. Install prerequisites (determine installation method based on Scrapling official docs and your OS)  
2. Register the Scrapling MCP Server, ensure it works after CLI restart  
3. Register the /research and /research-update commands  
After each step, confirm the result. After completion, read VERSION to confirm the version number and summarize the installation status.  

The AI will read the project docs → understand the system type → install each item → verify usability. No manual commands required.

🔧 Method 2: Non-OpenCode Users (Claude Code / Codex CLI / Cursor, etc.)

Paste this prompt into your AI coding tool:

Please research the project at https://github.com/hoolulu/deep-research, automatically install prerequisites and adapt it to the current CLI tool:  
1. Install Python and Scrapling (see Scrapling official docs and your system for method)  
2. Register the Scrapling MCP Server, restart to activate  
3. Register custom commands equivalent to /research and /research-update  
4. Translate the Task chain architecture into an equivalent implementation for the current tool  
After each step, confirm the result. After completion, read VERSION to confirm the version number and summarize the status.  

Adaptation points for different tools: multi-agent orchestration needs to map to each tool’s native mechanisms (Claude Code’s sub-agent, Codex CLI’s multi-file tasks, Cursor’s agent mode, etc.). The search and scraping logic (python-scrapling + search API) can be reused as-is.

Prerequisites

This skill depends on the sub-agent provided by the oh-my-openagent plugin. Without it, the /research command cannot execute. Other programming tools have their own multi-agent frameworks, so it may not be required.

X. Usage

After installation and restart of OpenCode, type in the chat:

What happens after you send it

The entire process runs automatically—you don’t need to do anything:

1 Analyze outline — analyze the topic, generate research framework and search plan  
2 Collect data — SearXNG / Exa cascading search → Scrapling batch scraping → data pool extraction → data quality check  
3 Parallel writing — multiple chapters written simultaneously, facts embedded directly in prompts, no additional tool calls  
4 Assemble and validate — batch validate → assemble-report → convert-citations → qa-report  

Total time ~10–20 minutes. Complex topics may take longer, simple ones may be faster.

Output file

The report is saved in Markdown format to the reports/ folder inside the skill directory, with a date-time stamp in the filename:

~/.opencode/skills/deep-research/reports/  

Open it with any Markdown reader (Typora / Obsidian / VS Code, etc.).

You can also specify a custom storage path by asking AI to modify it.

👉 Browse all generated reports — Click to browse all generated research reports; you can click to open and read.

XI. Cost

Estimates based on DeepSeek v4 Flash; actual figures may vary depending on model and topic complexity.

XII. FAQ

1. Search quota? How to ensure uninterrupted search?
The system uses a three-layer cascading search architecture. Each layer’s engine is independent; if the upper layer fails, it automatically falls back to the lower layer:

• Layer 1 — SearXNG (author deployed): A meta search engine deployed by the author on a VPS, aggregating 70+ search engines (including Baidu/Google/Brave), covering both Chinese and English. Built-in default endpoint, ready to use out of the box—unlimited, unthrottled, no quota.
• Layer 2 — Exa (cold backup): Built-in search engine in OpenCode, automatically configured by the OMO plugin, zero cost. If rate limited, automatically falls to Layer 3.
• Layer 3 — Free source reinforcement (final fallback): DuckDuckGo / Bing / Brave / Mojeek / Semantic Scholar / GDELT / arXiv + Baidu Baike / Zhihu / 199IT / iResearch / 36Kr / The Paper / East Money / Weibo / CSDN / Huxiu / Douban, 20+ sources. No API key required, always available.

2. How to generate a report using local materials?
You can try sending these three prompts to AI after installing the skill, adjusted to your needs. However, it’s best to first ensure your LLM can parse PDF and DOCX content; plain text and Markdown files can be read directly.

• Scenario 1: Local materials + online supplement (recommended for the most complete research)

Please use the deep-research skill. Based on the local materials in D:\My Notes\ProjectA, generate a research report about XX (quick mode). Prefer materials from the local files; if insufficient, supplement by searching online.  

• Scenario 2: Local materials only, no online search (when materials are sufficient and you worry online search might distract)

Please use the deep-research skill. Based on the local materials in D:\My Notes\ProjectA, generate a research report about XX (quick mode). Only use local materials; do not search online.  

• Scenario 3: Purely local, no skill, lightest (when you don’t need a professional report format, just a simple collation)

Based on the materials in D:\My Notes\ProjectA, help me organize them into a structured research report with a table of contents and chapter headings.  

3. How to update to the latest version?

Version strategy: The main branch always contains the latest code; minor changes are pushed directly. GitHub Releases are only used for milestone version markers (e.g., v2.1.0 → v2.2.0). There’s no need to wait for a new Release to update.

OpenCode users:

• Automatic: Type /research-update, AI will automatically run git pull to fetch the latest.
• Manual: cd ~/.opencode/skills/deep-research && git pull

Check the version number with cat ~/.opencode/skills/deep-research/VERSION.

4. Can non-OpenCode users update automatically?
You can simply ask AI to perform version comparison and update adaptation. Paste the following prompt into your AI coding tool:

Please compare the differences between the latest version of https://github.com/hoolulu/deep-research and your local version,  
identify new features and fixes from upstream, and apply them one by one to your local adapted version, preserving platform-specific modifications.  

5. Is the data secure?
All processing is done locally. No user data is collected or uploaded.

License

MIT

This project is licensed under the MIT License. MIT was chosen over stricter licenses (GPL, CC, etc.) because the project’s core is a portable methodology and pipeline design, not a copyrighted library of finished works. MIT allows it to be reused and modified across different platforms and toolchains to the greatest extent, consistent with the “not exclusive to OpenCode” positioning.

Created by hoolulu (https://github.com/hoolulu) · Project repository: github.com/hoolulu/deep-research (https://github.com/hoolulu/deep-research)

Community discussion: LINUX DO (https://linux.do/t/topic/2312664)