SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsmulti-ai-research

multi-ai-research

Writing & research

View repository

Updated 10-16 00:04

About this skill

Multi-AI Research (parallel multi-AI cross-validation)

Core value: Capability multiplication, not addition. Claude (main brain) + grok (X community/real-time) + gemini (Google ecosystem/structured) + N internal sub-agents = N+3 agents in parallelworking on the same research question.

Key insight(verified in practice on 2026-04-08): The consensus signal of two independent external AIs is stronger than the depth of any single AI. "More depth" < "fewer errors".

Relationship with ask-opencli：

  • ask-opencli = a single grok or gemini call (daily second opinion)
  • multi-ai-research = Full research workflow, parallel multi-AI + internal data + cross-validation + automatic arbitration

If the user just wants to "ask grok a question", use ask-opencli. If the user wants to do "deep research" or "cross-validate multiple dimensions", use this skill.

When to trigger

✅ Suitable

  • Deep research tasks (would take >30 minutes manually)
  • Industry mechanism questions (algorithm rules, product decisions, community consensus)
  • Internal data inference that needs external consensus weighting
  • Time-sensitive questions (grok has real-time X data, gemini has the latest web index)
  • Reversing existing assumptions (arbitration when data and theory conflict)
  • Feasibility research on new tools/new practices

❌ Not suitable

  • Pure code reasoning (a single Claude is enough)
  • Tasks that require deep project context (external AIs don't know your codebase)
  • Quick factual Q&A (solvable in <30 seconds, parallel overhead isn't worth it)
  • Creative generation (a single strong model is enough)

Workflow (7 Phase)

Phase 1: Problem decomposition (Claude automatic + user can override)

From one user research question, automatically decompose into:

  1. Internal data query(1-5 sub-agents in parallel)

    • Data distribution/quantitative analysis
    • Content comparison/qualitative analysis
    • Multi-dimensional segmentation
    • Time series/trend analysis
    • (add as needed)

  2. External theory query(2 Bash in parallel)

    • grok: focuses on real-time/community/X signals
    • gemini: focuses on structured/framework/long reasoning

Default decomposition strategy：

  • 3 internal agents + 2 external AIs = 5 parallel tasks
  • If the user's question is more theoretical → reduce internal agents to 1-2, increase the weight of external AIs
  • If the user's question is more data-oriented → increase to 4-5 internal agents, run only 2 external AIs for cross-checking

User can override: When the user explicitly says "only ask grok and gemini" or "only dispatch internal agents", follow the user's instruction.

Phase 2: Prompt automatic generation

For each parallel task, automatically generate a specific prompt:

Internal agent prompt template

你的任务是**只读数据分析**，不要修改任何文件。

## 背景
{{研究问题的 2-3 句背景描述}}

## 数据源
{{数据库路径或文件列表}}

## 任务
{{具体要查的维度，1-5 个 task}}

## 输出格式
- 结构化 markdown 报告
- 每个结论标注 n（样本数）和 置信度
- 3 屏幕内
- 纯文本返回，不要尝试写文件

grok / gemini prompt template

{{研究问题的精简描述，≤300 字}}

具体问：
(1) {{子问题 1}}
(2) {{子问题 2}}
...

请基于 2026 年上半年真实情况/最新数据回答，要具体可引用。

Key requirement: The prompts asked to grok and gemini must be consistent(premise of independent comparison).

Phase 3: Parallel dispatch (multiple tool calls in one message)

Tool 1: Agent (general-purpose)  run_in_background=true  [内部数据 agent A]
Tool 2: Agent (general-purpose)  run_in_background=true  [内部数据 agent B]
Tool 3: Agent (general-purpose)  run_in_background=true  [内部数据 agent C]
Tool 4: Bash run_in_background=true  [OPENCLI_BROWSER_COMMAND_TIMEOUT=300 opencli grok ask "..." --timeout 300 -f json]
Tool 5: Bash run_in_background=true  [opencli gemini ask "..." --format plain]

Must be in the same assistant message from ClaudeCall multiple tools at once to truly parallelize.

Phase 4: Wait (do not poll)

Background agent / Bash task completion willautomatically notify. While waiting, Claude can:

  • Read existing docs / memory to get more context
  • Prepare the output structure
  • Do a preliminary Phase 5/6 analysis framework

Explicitly prohibited：

  • Do not sleep + poll task status
  • Do not proactively Read the output file (unless a completion notification is received)
  • Do not jump ahead and draw conclusions

Phase 5: Cross-validation + automatic confidence grading (core of the major version)

After all results come back,according to the 4-tier classification rulesautomatically arbitrate each finding:

Classification rule matrix

Tier Judgment condition Confidence Handling
🟢 Strong consensus Internal data (n≥20) + grok + gemini all support High Can be written directly into the final conclusion as hard evidence
🟡 Partial consensus Internal data + 1 external AI support Medium Minor suggestion, retain skepticism
🔴 Conflict Internal data vs external AI contradiction Needs arbitration Decide according to arbitration principles
⚪ Insufficient data Internal data n<10 and no strong external support Low Mark as hypothesis, needs A/B testing

Arbitration principles (executed automatically)

Principle 1: Sample size threshold

  • n ≥ 20: can be used as high-confidence evidence
  • 10 ≤ n < 20: medium confidence, minor changes
  • n < 10: observation only, not used as a conclusion
  • n < 5: ignore completely

Principle 2: When data vs theory conflict

  • Internal data n ≥ 20 → prioritize internal data(unless both external AIs independently object)
  • Internal data n < 10 → prioritize external AI consensus(when both sides agree)
  • Internal data n < 5 → conclusion pending, mark as open question

Principle 3: Identification of suspicious concepts from external AI

  • gemini sometimes gives concepts suspected of being hallucinations (such as specific names like "Phoenix architecture")
  • Rule: concept nameappears only in gemini alone and has no grok cross-check → mark as "potential hallucination", do not adopt
  • Community data cited by grok ("users feedback says...") → medium credibility

Principle 4: Special handling of counterintuitive findings

  • If n ≥ 20 data contradicts common sense/theory (such as "small posts > big viral hits")
  • and external AI consensus supports it
  • Mark "⚠️ counterintuitive but strong consensus", needs follow-up verification

Phase 5.5: Secondary verification of external AI cases (new in the major version)

External AI sometimes gives specific "cases" (such as @morsyxbt 100→1600/月), such "single-source unique cases" need secondary verification:

Verification methods(by priority):

  1. Direct verification with user tools: If it's an X account → use twitter -c user <handle> to check whether it actually exists
  2. Ask another AI in reverse: Ask gemini "Have you heard of @morsyxbt?" to see whether there is cross-memory
  3. Web search verification: Have the user search manually or use an additional web tool

2026-04-09 Real-world case：

  • Grok uniquely mentioned @morsyxbt 100→1600/month
  • Gemini didn't mention it
  • Use twitter -c user morsyxbt to verify:Real account, 10.2k followers, verified
  • Conclusion: The account is real, but the specific figure "100→1600/month" still needs to be confirmed by morsyxbt's own tweet
  • Tier: Upgraded from "single-source suspicious" to "single-source credible case"

Identification of potential hallucinations：

  • If verification fails (account doesn't exist/numbers don't match) → mark as hallucination, remove from report
  • If partially verified (account exists but numbers unproven) → mark as partial verified, keep but lower the weight

Phase 6: Automatically generate a change list (tiered action items)

Output structure:

## Action Items (auto-generated, tiered)

### 🔴 极高置信度必做（Strong consensus，可直接落地）
1. [action] — 依据：{数据来源 + AI 共识} — 预期效果：{具体可衡量}
   - Rollback：{怎么撤销}
   - Verify：{24-48h 后如何验证}
2. ...

### 🟡 高置信度建议做（Partial consensus）
1. [action] — 依据：{单侧支持} — 前提假设：{什么条件下才成立}
2. ...

### ⚪ 待验证假设（Insufficient data）
1. [hypothesis] — 需要：{什么数据才能确认} — 建议：{A/B 测试设计}
2. ...

### 🚫 不做（Conflict / 反对证据强）
1. [originally planned action] — 反对依据：{为什么不做}

Phase 7: Artifact saving

Save to .omx/artifacts/multi-ai-research-<slug>-<YYYYMMDD-HHMMSS>.md

Must include:

  1. Original research question(the user's one sentence)
  2. Phase 1 question breakdown(the specific prompt for each task)
  3. Phase 3 parallel task list(task id + command)
  4. Raw response from each agent(N internal + 2 external)
  5. Phase 5 cross-verification matrix(confidence determination for each finding)
  6. Phase 6 change list（tiered action items）
  7. Time taken + number of tasks（metadata）

⚠️ Command quick reference (foolproof, look at this first)

The only valid subcommands for grok and gemini are these; everything else is wrong:

# ✅ Grok —— 只有一个子命令 ask
OPENCLI_BROWSER_COMMAND_TIMEOUT=300 opencli grok ask "问题" --timeout 300 -f json

# ✅ Gemini —— 5 个子命令，最常用是 ask
opencli gemini ask "问题" --format plain         # 最常用（单次问答）
opencli gemini new                               # 开新对话
opencli gemini deep-research "问题"              # Deep Research
opencli gemini deep-research-result              # 取 Deep Research 结果
opencli gemini image "画一个..."                 # 生图

❌ Common wrong commands (running them will directly report unknown command）：

❌ Wrong ✅ Correct Note
opencli gemini chat "..." opencli gemini ask "..." There is no chat subcommand(common mistake, don't use it out of habit)
opencli gemini query "..." opencli gemini ask "..." There is no query
opencli grok chat "..." opencli grok ask "..." grok only has ask
opencli grok new opencli grok ask "..." --new true grok's "new conversation" is a parameter of ask

Memory anchor：ask is the only "ask once" command for both. Not chat, not query, not prompt。

If unsure, run opencli grok --help or opencli gemini --help to see the full subcommand list.

First-time use: Install opencli (5-minute one-time setup)

Official repository：https://github.com/jackwener/opencli npm package：@jackwener/opencli (npmjs) Author：jackwener License：Apache-2.0

Environment requirements

  • Node.js ≥ 20.0.0
  • Chrome or Chromium browser
  • Supported on macOS / Linux / Windows

1. Install the opencli CLI

npm install -g @jackwener/opencli

Verify:

opencli --version

2. Install the Browser Bridge extension

opencli uses alightweight Chrome extension + local daemon to reuse the session already logged in in your browser. On first run it will automatically guide you through installation:

opencli doctor

Load the extension into Chrome as prompted (usually chrome://extensions → Developer mode → Load unpacked, and the doctor will tell you the path).

3. Log in to the target AI website

Open and log in using the Chrome profile with the extension installed:

Logging in once is enough; the session will be reused by opencli long-term.

4. Set environment variables (required)

# 加到 ~/.zshrc 或 ~/.bashrc
export OPENCLI_BROWSER_COMMAND_TIMEOUT=300

Why it is required: opencli's default browser command timeout is 60 seconds (runtime.js:25), which is not enough for complex grok questions. If you do not set this, grok will report timed out after 60s. This is a lesson learned the hard way.

5. Optional: install opencli's AI skills

opencli itself provides several AI skills, which you can also install:

npx skills add jackwener/opencli

(These skills and multi-ai-research do not conflict; they are complementary.)

6. Verify the full chain

OPENCLI_BROWSER_COMMAND_TIMEOUT=300 opencli grok ask "请只回复：OK" --timeout 300 -f json
opencli gemini ask "请只回复：OK" --format plain

If both return OK, that means the full chain is working.

Extended Reference

Detailed material starting at ## Phase 0: Pre-flight Prerequisites（强制检查） has been moved to reference/extended.md to keep this skill concise. Load that reference when the task requires the moved examples, command catalogs, checklists, platform details, or implementation templates.