tech-research

中文文档

A Claude Code skill for multi-source technical research. Gathers intelligence from three sources and synthesizes findings into a unified report — using lightweight subagents for simple queries or coordinated agent teammates for heavy competitive research.

Source	Tool	What It Provides
Grok	Browser automation → grok.com	X (Twitter) developer discussions, sentiment, expert discovery
DeepWiki	DeepWiki MCP → ask_question	AI-powered GitHub repository analysis, architecture, API docs
WebSearch	Built-in web search	Official docs, benchmarks, blog posts, recent announcements

Installation

Via skills.sh (recommended)

npx skills add psylch/tech-research-skill -g -y

Via Claude Code Plugin Marketplace

/plugin marketplace add psylch/tech-research-skill
/plugin install tech-research@psylch-tech-research-skill

Restart Claude Code after installation.

Prerequisites

• Claude Code or any agent that supports skills.sh
• Browser automation for Grok (optional — Grok source is skipped if unavailable):
  • Claude-in-Chrome (zero setup, recommended), or
  • Playwright MCP (auto-setup via grok_setup.sh)
• DeepWiki MCP for GitHub repository analysis (optional)

Grok Browser Backend

Grok requires browser automation with a logged-in session. The skill supports multiple backends, detected in priority order:

Priority	Backend	MCP Server	Setup
1	Claude-in-Chrome	claude-in-chrome	Zero setup — uses your Chrome login state
2	Playwright-Grok	playwright-grok	One-time: run grok_setup.sh setup to create from existing Playwright config
3	Playwright (default)	playwright	Works but no login persistence

Important: Do NOT modify your default playwright MCP to add --user-data-dir. This forces ALL browser operations through a single profile, breaking parallel agent usage. The skill uses a separate playwright-grok instance instead.

Quick Setup (if no Claude-in-Chrome)

If you have Playwright MCP configured, run:

bash ~/.agents/skills/tech-research/scripts/grok_setup.sh setup

This clones your playwright config as playwright-grok with a dedicated profile directory. Then restart Claude Code and log into Grok once in the new browser window.

Login Status Cache

Login state is cached at ~/.claude/tech-research/.grok-status.json:

• logged_in — Long-lived, no expiry. Valid until a subagent detects an actual logout.
• logged_out — Auto-expires after 2 hours, then retries optimistically.
• No file — Optimistic, assumes logged in.

# Check current status
bash ~/.agents/skills/tech-research/scripts/grok_setup.sh check

# Clear cached status (after re-logging in)
bash ~/.agents/skills/tech-research/scripts/grok_setup.sh reset

Usage

In Claude Code, use any of these trigger phrases:

/tech-research
调研一下 Zustand vs Jotai
research this technology: Bun
compare libraries: Vite vs Turbopack
evaluate framework: SolidJS

How It Works

1. Analyze — breaks the research question into sub-queries for each source
2. Pre-flight — detects best available browser backend and login status for Grok
3. Select mode — chooses Light or Heavy mode based on research complexity
4. Dispatch — launches research agents in the appropriate mode
5. Synthesize — merges findings into a structured report with TL;DR, comparison matrix, and actionable recommendation

Research Modes

Signal	Mode
Single topic, multiple data sources	Light — up to 3 parallel Task Subagents, each handling one data source
Multiple topics/competitors needing cross-comparison	Heavy — Agent Teammates that can communicate, share discoveries, and avoid duplication
Research may require dynamic re-scoping	Heavy
Agent count ≥ 4	Heavy

Not every research task uses all 3 sources. The skill selects sources based on the question type:

Research Type	Grok	DeepWiki	WebSearch
"Should we use library X?"	Yes	Yes	Yes
"What are devs saying about X?"	Yes	No	Maybe
"How does repo X work internally?"	No	Yes	Maybe
"Compare X vs Y performance"	Maybe	Yes (both)	Yes

Key Design Decisions

• Grok queries must include X/Twitter keywords (e.g., "developers on X say...") to avoid falling back to broad web search, which would duplicate WebSearch
• Each Grok query opens a fresh page — no multi-turn conversations in the same session
• DeepWiki uses ask_question only — read_wiki_structure and read_wiki_contents return massive dumps that easily exceed context limits
• Grok post URLs are verified — the subagent navigates to 2-3 cited URLs to check if they actually exist
• Optimistic login — assumes Grok is logged in unless a previous attempt recorded logged_out; avoids prompting the user unnecessarily
• Self-recovery — the skill includes a diagnostic decision tree for handling script failures, stale status files, and browser issues

Troubleshooting

Symptom	Fix
Grok always skipped despite being logged in	Run grok_setup.sh reset to clear stale status
grok_setup.sh check returns unexpected results	Verify ~/.claude.json has correct MCP server names
Status file corrupted	Delete ~/.claude/tech-research/.grok-status.json
Script fails entirely	The skill falls back to ToolSearch-based MCP detection automatically

File Structure

tech-research-skill/
├── .claude-plugin/
│   └── plugin.json                   # Plugin manifest
├── skills/
│   └── tech-research/
│       ├── SKILL.md                  # Main skill definition
│       ├── references/
│       │   ├── subagent_templates.md # Prompt templates for each subagent
│       │   └── query_strategies.md   # Grok query crafting strategies
│       └── scripts/
│           └── grok_setup.sh        # Backend detection, setup, and login status management
├── README.md
├── README.zh.md
└── LICENSE

License

MIT