tech-research（技术研究）

Tech Research

Orchestrate multi-source technical research by dispatching parallel subagents to gather intelligence from X/Twitter (via Grok), GitHub repositories (via DeepWiki), and the web (via WebSearch). Synthesize all findings into a single actionable report.

Architecture: The main agent orchestrates research using one of two modes — lightweight (Task Subagents) or heavyweight (Agent Teammates) — chosen based on research complexity.

Language

Match user's language: Respond in the same language the user uses. If the user writes in Chinese, the entire research report should be in Chinese. If in English, report in English.

Research Mode Selection

Before dispatching any agents, determine the appropriate mode:

Light Mode (default for single-topic research)

Dispatch up to 3 Task Subagents (Task with subagent_type: "general-purpose"). Each handles one data source independently. The main agent synthesizes results after all return.

Heavy Mode (for multi-topic / competitive research)

Use TeamCreate to create a research team → TaskCreate for each research task → spawn Agent Teammates (via Task with team_name and name parameters) → coordinate via SendMessage. Teammates can:

• Communicate to avoid duplication ("I found Project A uses the same approach as B — focus on their differentiators")
• Share discoveries across tasks ("The blog post I found compares all 3 frameworks, sending you the link")
• Dynamically adjust scope based on what others have found

When to Use

• Evaluating a technology, library, or framework for adoption
• Comparing alternatives (e.g., "Zustand vs Jotai vs Redux")
• Investigating a GitHub repo's architecture and community reception
• Gauging developer sentiment on a new API, tool, or announcement
• Cross-language research (Chinese/Japanese developer communities)

Research Sources

Source Degradation

Not every source will be available every time. Follow this degradation strategy:

Minimum viable research: At least one source must return results. If all sources fail, report the failures and suggest the user check their environment setup.

Grok Browser Backend

Grok requires browser automation with login state. Multiple backends are supported, detected in priority order:

Detection is strict first-match, top-down. As soon as a priority's detector returns true, set backend to that row and stop. Do NOT check lower priorities. CLI backends (0 and 1) skip grok_setup.sh check entirely — it only tracks MCP config state, not CLI presence.

IMPORTANT: Do NOT modify the user's default playwright MCP to add --user-data-dir. This would force ALL browser operations through a single profile, breaking parallel agent usage. Instead, use a separate playwright-grok instance.

Grok Pre-flight

Before dispatching a Grok subagent, walk the priority table top-down. Run each detector; on first match, set backend and stop. Paths are mutually exclusive — never evaluate a lower priority after a higher one has matched.

Priority 0 — better-agent-browser (CLI):

command -v agent-browser >/dev/null && echo "MATCH: backend=better-agent-browser"

If matched: set backend=better-agent-browser. The Grok subagent loads the better-agent-browser skill and follows its Layer 0b recipe (headless, dedicated ~/.chrome-debug-profile on port 9333). Stop. Do not check lower priorities.

Priority 1 — browser-use (CLI):

command -v browser-use >/dev/null && echo "MATCH: backend=browser-use"

Only evaluated if priority 0 did NOT match. If matched: set backend=browser-use. The Grok subagent drives browser-use directly per its own docs. Stop.

Priorities 0 and 1 do NOT depend on ~/.claude.json or MCP runtime availability. Do NOT run grok_setup.sh check on these paths.

Priority 2 — claude-in-chrome (runtime MCP): Only evaluated if priorities 0 and 1 did NOT match. Check whether mcp__claude-in-chrome__* tools are visible in your current session. If yes: set backend=chrome. The chrome extension is injected at runtime and never appears in ~/.claude.json, so the shell script cannot detect it. Skip the preflight script entirely. Stop.

Priority 3–4 — Playwright fallback (script-based detection): Only evaluated if priorities 0–2 did NOT match. Run the preflight script to detect playwright backends:

bash ${SKILL_PATH}/scripts/grok_setup.sh check

The script outputs preflight JSON: {"ready": true, "backend": "playwright-grok", "login_status": "logged_in", "hint": "..."}.

Key fields: ready (boolean), backend (playwright-grok/playwright/none), login_status (logged_in/logged_out/unknown), hint (human-readable summary).

Live validation note: Preflight checks MCP server presence in ~/.claude.json (existence check) but cannot live-test browser MCP connectivity from shell — MCP servers are runtime-managed by Claude Code and only accessible during agent execution. The skill compensates with optimistic dispatch: assume the backend works, then update the login status cache based on actual browser interaction results (see Login Status Cache below).

Config safety: The setup command always creates a timestamped backup of ~/.claude.json before writing. If the new config causes issues, restore with the rollback command in the setup output (e.g., cp ~/.claude.json.backup.<timestamp> ~/.claude.json).

Login Status Cache

Login state is cached at ~/.claude/tech-research/.grok-status.json. Cache semantics: logged_in persists until a subagent observes logout; logged_out auto-expires after 2 hours (or clear with grok_setup.sh reset); unknown/missing = optimistic, try Grok.

Subagent responsibility: After interacting with Grok, update login status:

• Success: bash ${SKILL_PATH}/scripts/grok_setup.sh status logged_in <backend>
• "Sign in" page: bash ${SKILL_PATH}/scripts/grok_setup.sh status logged_out <backend>

Cache corruption recovery: If the status file is corrupted (invalid JSON, permission errors), delete it (rm ~/.claude/tech-research/.grok-status.json) and proceed in optimistic mode. See references/troubleshooting.md for the full decision tree.

Execution Context: Top-Level vs Subagent

This skill is a two-layer contract:

• SKILL.md (this file) — orchestrator's manual. Covers preflight, mode selection, dispatch, and synthesis. This is what the agent running tech-research directly should read.
• references/subagent_templates.md — per-role contracts (Grok / DeepWiki / WebSearch). Each template is complete and self-contained — a dispatched role-agent does NOT need to read SKILL.md. The orchestrator points the role-agent to the right section by path and passes minimal parameters.

Before starting, determine how this skill was invoked:

How to detect: if Task is not in your available tools (or ToolSearch select:Task returns no match), you're in Inline Mode.

Either way, the per-role behavior (login check, query rules, report format, deferred-tool loading) lives in the template file, not here. Both modes converge on the same reference contract.

Workflow

Progress:

• Step 1: Preflight — Run grok_setup.sh check, complete any setup BEFORE research
• Step 2: Analyze — Break question into per-source sub-queries
• Step 3: Dispatch — Launch subagents (Light/Heavy Mode) OR run inline (Inline Mode)
• Step 4: Synthesize — Merge findings into unified report

1. Preflight Gate (MUST run first)

Run preflight before doing anything else. This determines which sources are available and whether one-time setup is needed.

Walk priorities top-down; stop on first match.

1. command -v agent-browser succeeds → backend=better-agent-browser. Stop, go to Step 2 of workflow.
2. Otherwise, command -v browser-use succeeds → backend=browser-use. Stop, go to Step 2 of workflow.
3. Otherwise, mcp__claude-in-chrome__* tools visible in session → backend=chrome. Stop, go to Step 2 of workflow.
4. Otherwise (priorities 0–3 all missed), run bash ${SKILL_PATH}/scripts/grok_setup.sh check to probe for playwright MCP backends. Only reach this branch if none of the above matched — if you matched priority 0/1/2, you already stopped.

Exit code 0 (READY)     → Note the backend and login_status, proceed to Step 2.
Exit code 1 (NEEDS_SETUP) → Run `grok_setup.sh setup` NOW.
                            If setup succeeds, tell the user:
                              "playwright-grok has been configured. Restart Claude Code
                               to activate it. For this session, Grok will use the
                               default playwright backend as a fallback."
                            Then proceed to Step 2 with backend=playwright.
Exit code 2 (NOT_AVAILABLE) → Grok is unavailable. Proceed to Step 2 without Grok.

Key principle: Setup completes before research begins. Never discover setup needs mid-research.

2. Analyze the Research Question

Break the user's question into sub-queries for each source:

• Grok query: Developer opinions, community sentiment, expert recommendations
• DeepWiki query: Repository architecture, API design, code quality (requires owner/repo)
• WebSearch query: Official docs, benchmarks, comparisons, recent announcements

Not every research task needs all 3 sources. Select sources based on the question:

3. Dispatch Research Agents

Choose the dispatch method based on the research mode and execution context.

Browser is a singleton — enforce it in the prompts. Only the Grok role touches a browser. The DeepWiki and WebSearch role prompts MUST forbid loading any browser skill (better-agent-browser, playwright, claude-in-chrome) — they are restricted to DeepWiki MCP, WebSearch, and WebFetch only. Otherwise parallel roles race for the same Chrome user-data-dir / CDP port and collide.

If you genuinely need multiple parallel browser roles (rare), switch to Heavy Mode and coordinate all of them through one shared Layer-2 CDP proxy — see better-agent-browser Layer 2.

Light Mode: Task Subagents (top-level only)

Each role-agent is pointed at its contract by path. Do NOT inline the template body into the Task prompt — the subagent reads the file directly. This keeps the dispatch prompt small and guarantees the subagent sees the canonical, up-to-date contract.

Grok role:

Task(
  subagent_type: "general-purpose",
  description: "Grok research [topic]",
  prompt: """
  Read ${SKILL_PATH}/references/subagent_templates.md § Grok Subagent Template.
  That file is your complete contract — do NOT read SKILL.md.

  Parameters:
  - RESEARCH_QUESTION: [question]
  - BACKEND: [backend from preflight]
  - GROK_QUERY: [X/Twitter-scoped query per the template's rules]

  Execute literally and return findings in the format the template specifies.
  """
)

DeepWiki role:

Task(
  subagent_type: "general-purpose",
  description: "DeepWiki research [repo]",
  prompt: """
  Read ${SKILL_PATH}/references/subagent_templates.md § DeepWiki Subagent Template.
  That file is your complete contract — do NOT read SKILL.md.
  Do NOT load any browser skill.

  Parameters:
  - RESEARCH_QUESTION: [question]
  - REPO_LIST: [owner/repo, ...]
  - CUSTOM_QUESTIONS: [topic-specific questions]

  Execute literally and return findings in the format the template specifies.
  """
)

WebSearch role:

Task(
  subagent_type: "general-purpose",
  description: "Web research [topic]",
  prompt: """
  Read ${SKILL_PATH}/references/subagent_templates.md § WebSearch Subagent Template.
  That file is your complete contract — do NOT read SKILL.md.
  Do NOT load any browser skill.

  Parameters:
  - RESEARCH_QUESTION: [question]
  - TOPIC: [topic keywords]
  - CURRENT_YEAR: [year]

  Execute literally and return findings in the format the template specifies.
  """
)

Inline Mode (running inside an existing subagent)

You cannot spawn further subagents. Instead, read references/subagent_templates.md yourself and execute each role's contract in your own tool loop:

1. Read the template file once.
2. For each role you need, treat the template section as your instructions and execute inline.
3. Run WebSearch + DeepWiki in parallel tool calls. Run Grok separately (it holds the browser singleton).
4. Honor the browser-singleton rule: don't try to run multiple browser-using steps concurrently.
5. Collect all three role reports mentally, then proceed to Step 4 (Synthesize) as usual.

Heavy Mode: Agent Teammates

1. TeamCreate(team_name: "research-[topic]")
2. TaskCreate(subject: "Research [Project A]", description: "...", activeForm: "Researching [Project A]")
3. TaskCreate(subject: "Research [Project B]", description: "...", activeForm: "Researching [Project B]")
4. Task(subagent_type: "general-purpose", team_name: "research-[topic]", name: "researcher-a", prompt: "...")
5. Task(subagent_type: "general-purpose", team_name: "research-[topic]", name: "researcher-b", prompt: "...")
6. Coordinate via SendMessage — share findings, adjust scope, avoid duplication
7. Synthesize after all teammates report back
8. Shutdown teammates and TeamDelete when done

Each teammate should use all relevant data sources (Grok, DeepWiki, WebSearch) for their assigned topic, rather than splitting by data source.

4. Synthesize and Report

After all subagents return, merge findings into a unified report with these sections:

• TL;DR — 2-3 sentence executive summary with clear recommendation
• Community Sentiment (from X/Twitter) — Key opinions with @username attribution and post URLs
• Repository Analysis (from DeepWiki) — Architecture, code quality, API design, maintenance status
• Web Intelligence — Official docs, benchmarks, blog insights, announcements
• Comparison Matrix (if comparing alternatives) — Criteria-based table
• Recommendation — Clear, actionable recommendation based on all sources
• Limitations — What couldn't be verified, including skipped sources and why

Grok Query Strategies

See references/grok_query_overview.md for strategy selection table and references/query_strategies.md for full prompt templates.

Troubleshooting & Self-Recovery

See references/troubleshooting.md for the full decision tree and common scenarios.

Tips

• For CJK communities, query Grok in the target language directly
• DeepWiki accepts up to 10 repos in a single query for comparisons
• WebSearch is best for recent information (include current year in queries)
• Always verify Grok post URLs before citing — accuracy is ~80%
• Run subagents in parallel to minimize total research time