Multi-agent web research with mandatory URL verification, confidence-tagged output, and four depth modes (quick to deep investigation). USE WHEN research, do re
复制下面这句话，粘贴给 Claude Code、Codex、Cursor 等 AI 编程工具，它会读取安装说明并在你确认后完成安装。
请阅读 https://ai.atlankj.com/install/asset/gh-research-71d8d85115d5 ，按照其中的说明把「Research」安装到你（当前 AI 工具）中。执行前先告诉我将运行的命令和写入的位置，等我确认。
查看 AI 将读取的安装说明正在读取 GitHub 原文…
内容来自 GitHub 原始文件，由原作者维护。在 GitHub 查看
When user says "research" (in any form), ALWAYS invoke this skill.
| User Says | Action |
|---|---|
| "research" / "do research" / "research this" | → Standard mode (3 agents: Claude + Gemini + Perplexity + cross-check) |
| "quick research" / "minor research" | → Quick mode (1 Perplexity agent) |
| "extensive research" / "deep research" | → Extensive mode (7 explorers + 2 verifiers) |
| "deep investigation" / "investigate [topic]" / "map the [X] landscape" | → Deep Investigation (iterative + verification) |
"Research" alone = Standard mode. No exceptions.
Deterministic alternative (EXPERIMENTAL — not yet run in the harness): Workflows/research.mjs ports Standard + Extensive into a Workflow-tool script — fixed researcher roster, single batch URL-verify, cross-checked synthesis. It is parse-verified and contract-checked but has NOT yet had a live harness run, so the prose StandardResearch.md / ExtensiveResearch.md stay the default path. Do not route real research through the .mjs until one smoke run lands. To do that smoke run: Workflow({ scriptPath: "skills/Research/Workflows/research.mjs", args: { question: "<trivial test>", depth: "standard" } }). Once it runs clean, drop this experimental caveat.
Before executing, check for user customizations at:
~/.claude/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Research/
If this directory exists, load and apply any PREFERENCES.md, configurations, or resources found there. These override default behavior. If the directory does not exist, proceed with skill defaults.
You MUST send this notification BEFORE doing anything else when this skill is invoked.
Send voice notification:
curl -s -X POST http://localhost:31337/notify \
-H "Content-Type: application/json" \
-d '{"message": "Running the WORKFLOWNAME workflow in the Research skill to ACTION"}' \
> /dev/null 2>&1 &
Output text notification:
Running the **WorkflowName** workflow in the **Research** skill to ACTION...
This is not optional. Execute this curl command immediately upon skill invocation.
Researches a topic across multiple sources and verifies every claim before delivery. Four depth modes scale from a single fast lookup to a multi-session investigation: Quick (1 agent, ~10-15s), Standard (3 agents cross-checked, ~30-60s), Extensive (7 explorers + 2 independent verifiers, ~60-90s), and Deep Investigation (progressive iteration with a persistent vault, ~3-60min). Output is confidence-tagged: [HIGH] [MED] [LOW] [CONFLICT].
A single AI agent doing research has two failure modes that quietly wreck the result. It hallucinates URLs — confident links that go nowhere, which destroys trust in the whole report. And it answers from one angle, so it parrots whatever the first few search results said and misses conflicts, gaps, and what real people actually thought. Recap journalism is the worst offender: ask "what did fans think of X" and a lone agent hands back promoter copy dressed as consensus. This skill runs several agents in parallel, cross-checks and independently verifies their findings, checks every URL before it ships, and routes sentiment questions to community sources first.
Multiple agents work in parallel and their findings get reconciled. Verification runs in three layers at zero added latency: each agent self-verifies its own URLs, a synthesis step cross-checks for conflicts, and dedicated verifier agents (Extensive/Deep) check findings with no access to the explorers' reasoning. Step 0 of every workflow routes sentiment questions to community scrapers before web search, and every URL is verified before delivery — a hallucinated link is a catastrophic failure.
Official anchor (drift check): this skill's orchestrator-worker shape matches Anthropic's published architecture — How we built our multi-agent research system (2025-06-13). When changing the orchestration itself (rosters, verifier layers, the .mjs ports — not on routine research runs), re-fetch that post and compare: it names lessons we may not have adopted yet (e.g. async agent coordination). Advisory only — report divergence, never auto-adopt; an unreachable URL never blocks anything.
READ: UrlVerificationProtocol.md - Every URL must be verified before delivery.
Research agents hallucinate URLs. A single broken link is a catastrophic failure.
READ: SourceRoutingProtocol.md — sentiment-signal detection + scraper-first paths for Reddit / YouTube / X / TikTok.
The rule: web search answers "what was published about X." Community scrapers answer "what people said about X." If the question is about fan sentiment, ratings, reactions, opinions, or what real people thought — route to Reddit (JSON API first, Apify fallback), YouTube comments, and X before spawning Perplexity/Claude/Gemini web-search agents. Recap journalism is the secondary source, not the primary one.
Sentiment signal triggers (run at Step 0 of Quick / Standard / Extensive):
Detection fires → sentiment-mode routing per SourceRoutingProtocol.md. Detection does not fire → standard routing.
Before executing any workflow, verify context sufficiency: do I have what I need to produce a hard-to-vary research artifact, or am I about to speculate? If the question shape and target sources are clear, proceed. If speculating, emit a one-line ambiguity flag and ship best-effort. If clearly insufficient, emit ≤3 questions with proceed override.
CRITICAL: For due diligence, company/person background checks, or vetting -> use a dedicated OSINT/entity-investigation skill instead
| Workflow | Trigger | File |
|---|---|---|
| QuickResearch | Quick/minor research; Perplexity API research (1 Perplexity agent, 1 query) | Workflows/QuickResearch.md |
| StandardResearch | Standard research — DEFAULT (3 agents: Claude + Gemini + Perplexity, cross-checked) | Workflows/StandardResearch.md |
| ExtensiveResearch | Extensive research (7 explorers + 2 verifiers = 9 agents) | Workflows/ExtensiveResearch.md |
| DeepInvestigation | Deep investigation / iterative research / map the [X] landscape (progressive deepening, loop-compatible) | Workflows/DeepInvestigation.md |
| DeepVerifiedResearch | Deep verified / fact-checked research — slowest tier, claim-level adversarial verification (see notes below) | Workflows/DeepVerifiedResearch.mjs |
| research.mjs | EXPERIMENTAL deterministic port of Standard + Extensive — do NOT route real research here until a smoke run lands (see Mandatory Trigger note) | Workflows/research.mjs |
| Verify | Verify research findings / cross-check claims / confidence scoring | Workflows/Verify.md |
| ExtractAlpha | Extract alpha / deep analysis / highest-alpha insights | Workflows/ExtractAlpha.md |
| Retrieve | Difficulty accessing content (CAPTCHA, bot detection, blocking) | Workflows/Retrieve.md |
| YoutubeExtraction | YouTube URL extraction (use fabric -y URL immediately) | Workflows/YoutubeExtraction.md |
| WebScraping | Web scraping | Workflows/WebScraping.md |
| ClaudeResearch | Claude WebSearch only (free, no API keys) | Workflows/ClaudeResearch.md |
| InterviewResearch | Interview preparation (Tyler Cowen style) | Workflows/InterviewResearch.md |
| AnalyzeAiTrends | AI trends analysis | Workflows/AnalyzeAiTrends.md |
| Fabric | Use Fabric patterns (242+ specialized prompts) | Workflows/Fabric.md |
DeepVerifiedResearch notes: run via Workflow({scriptPath: 'skills/Research/Workflows/DeepVerifiedResearch.mjs', args: {question: '...'}}) (pass args as an OBJECT, never a JSON string). Does NOT replace Extensive — it sits below it. Measured ~150-190s vs Extensive's ~60-90s, because claim-level verification needs one extra serial hop (you can't vote on claims until they're extracted). Reach for it only when claims must be bulletproof: each extracted claim is attacked by three skeptics from different lenses (quote-support, contradiction, source-strength), survives only on a quorum of non-refuting votes (all-abstain never survives), then a written synthesis frames the survivors with [HIGH]/[MED]/[LOW]/[CONFLICT] tags and refuted-claim transparency. Dedup, ranking, vote-counting, and the abstention guard run deterministically in the script. research.mjs is the faster sibling and the place multi-vendor diversity lives (Standard/Extensive rosters + URL verify).
READ: QuickReference.md for detailed examples and mode comparison.
| Trigger | Mode | Speed |
|---|---|---|
| "quick research" | 1 Perplexity agent | ~10-15s |
| "do research" | 3 agents + cross-check | ~30-60s |
| "extensive research" | 7 explorers + 2 verifiers | ~60-90s |
| "deep investigation" | Progressive iteration + verification | ~3-60min |
Inspired by Nomad (arXiv:2603.29353). Three layers of verification, zero added latency:
| Layer | What | Where | Cost |
|---|---|---|---|
| Self-Verification | Each agent verifies own URLs and tags confidence before returning | All agents | 0s (inside parallel window) |
| Cross-Check | Synthesis step detects conflicts and cross-references findings | Standard, Extensive, Deep | 2-3s (within synthesis) |
| Independent Verification | Dedicated verifier agents with no access to explorer reasoning | Extensive, Deep only | 0s (parallel with explorers) |
Confidence tags in output: [HIGH] [MED] [LOW] [CONFLICT]
See Workflows/Verify.md for full verification protocol.
Progressive iterative research that builds a persistent knowledge vault. Works in both single-run (one cycle) and iterative mode (Algorithm-driven iterations).
Concept: Broad landscape → discover entities → score importance/effort → deep-dive one at a time → loop until coverage complete.
Domain template packs customize the investigation for specific domains:
Templates/MarketResearch.md — Companies, Products, People, Technologies, Trends, InvestorsTemplates/ThreatLandscape.md — Threat Actors, Campaigns, TTPs, Vulnerabilities, Tools, DefendersExample invocation:
"Do a deep investigation of the AI agent market"
→ Loads MarketResearch.md template
→ Iteration 1: Broad landscape + first entity deep-dive
→ Iterative mode: Each iteration deep-dives the next highest-priority entity
→ Exit: When all CRITICAL/HIGH entities researched + all categories covered
Artifacts persist at ~/.claude/LIFEOS/MEMORY/RESEARCH/{date}_{topic}/ — the vault survives across sessions.
See Workflows/DeepInvestigation.md for full workflow details.
Working files (temporary work artifacts): ~/.claude/LIFEOS/MEMORY/WORK/{current_work}/
~/.claude/ to get the work_dir valueHistory (permanent): ~/.claude/History/research/YYYY-MM/YYYY-MM-DD_[topic]/
x\.com|twitter\.com (e.g. rg -q 'x\.com|twitter\.com' on the request text) BEFORE spawning any research agents. X blocks WebFetch and generic scraping, so generic research agents burn turns and return nothing. When the gate fires, do NOT spawn generic agents at the URL — take the first path below that is actually available, and say in the response which path you took and why (a silent skip is a failure):
X_BEARER_TOKEN in the environment → read the post via X API v2 directly.skills/Apify/skills/get-user-tweets.ts) or the BrightData ladder, if either is configured.SourceRoutingProtocol.md. Recap web search is the secondary source for community-sentiment questions, not the primary one. Quick mode can return recap-only and miss the actual fan data — pull Reddit directly rather than waiting to be asked again. Do not repeat.X_BEARER_TOKEN is set, YouTube Data API v3 if YOUTUBE_API_KEY is set) before reaching for Apify or BrightData. Scrapers are fallback for when the API path is unavailable, rate-limited, or doesn't expose the data shape needed (e.g., YouTube transcripts — use fabric -y even when the Data API key is set). The cascade inversion is the recurring failure mode. See SourceRoutingProtocol.md Cascade Priority section for the per-platform table..json to any thread or listing URL. Set User-Agent: LifeOS-Research/1.0 or Reddit rate-limits the default UA. Apify Reddit scraper is Tier 2 (fallback), not Tier 1.Example 1: Quick lookup
User: "quick research on Hono SSR middleware patterns"
→ Invokes QuickResearch workflow (1 Claude agent)
→ Returns summary with key patterns and links
→ ~10-15 seconds
Example 2: Standard multi-source research
User: "research the current state of AI agent frameworks"
→ Invokes StandardResearch workflow (3 agents: Claude + Gemini + Perplexity, cross-checked)
→ Cross-references findings, confidence-tags, verifies URLs
→ Returns synthesized report with citations
→ ~30-60 seconds
Example 3: Deep investigation
User: "do a deep investigation of the AI agent market"
→ Invokes DeepInvestigation workflow
→ Broad landscape scan → entity discovery → priority scoring → deep-dives
→ Builds persistent knowledge vault in MEMORY/RESEARCH/
→ Loop-compatible for multi-session investigation
After completing any workflow, append a single JSONL entry:
echo '{"ts":"'$(date -u +%Y-%m-%dT%H:%M:%SZ)'","skill":"Research","workflow":"WORKFLOW_USED","input":"8_WORD_SUMMARY","status":"ok|error","duration_s":SECONDS}' >> ~/.claude/LIFEOS/MEMORY/SKILLS/execution.jsonl
Replace WORKFLOW_USED with the workflow executed, 8_WORD_SUMMARY with a brief input description, and SECONDS with approximate wall-clock time. Log status: "error" if the workflow failed.
| Enhance | Enhance/improve content | Workflows/Enhance.md |
| ExtractKnowledge | Extract knowledge from content | Workflows/ExtractKnowledge.md |
fabric -y URL directly — don't try to scrape YouTube pages with WebFetch.DeepVerifiedResearch.mjs is a Workflow-tool script, not a markdown workflow. Invoke it with the Workflow tool (scriptPath), never by reading it and "doing the steps" — the whole point is that dedup, fetch-budget, vote-counting, and the abstention guard run deterministically in code. Running it spawns many agents + live web calls, so it is opt-in multi-agent: confirm with the principal (or use args.test: true for a small smoke run) rather than firing a full ~30–95-agent run unprompted.research.mjs (text-returning researchers), not in the schema-gated verification engine. A claim only survives a quorum of valid votes with fewer than the kill threshold refuting; all-abstain does NOT survive (guards the false-survive bug).