consultant（consultant 搜索）

Consultant Skill

Three Operation Modes

CONSULTATION MODE (Single Expert)

Ask a specific AI model for focused analysis or second opinion.

Trigger Patterns:

• "Ask [model] about X"
• "Consult [model] on Y"
• "What does [model] think about Z"
• "Get [model]'s opinion on X"
• "[Model], analyze this problem"

Available Consultants:

• codex / gpt5 - OpenAI GPT-5 Codex (advanced reasoning, technical analysis)
• gemini - Google Gemini (multi-perspective analysis)
• grok - xAI Grok (alternative LLM perspective)
• perplexity - Perplexity Sonar (web research with citations)
• claude - Claude WebSearch (web research, detailed analysis)

Examples:

• "Ask Gemini about the best approach to implement this feature"
• "Consult Codex on this architectural decision"
• "What does Grok think about this code pattern?"
• "Get Perplexity's research on latest React trends"

RESEARCH MODE (Multi-Agent Parallel)

Launch multiple agents simultaneously for comprehensive coverage.

Trigger Patterns:

• "Do research on X"
• "Quick research: X"
• "Extensive research on X"
• "Research this topic"
• "Investigate X"

Three Research Intensities:

QUICK RESEARCH

• Launch 5 agents (1 of each type)
• Timeout: 2 minutes
• Best for: Simple queries, fast answers

STANDARD RESEARCH (Default)

• Launch 15 agents (3 of each type)
• Timeout: 3 minutes
• Best for: Comprehensive coverage, balanced depth

EXTENSIVE RESEARCH

• Launch 40 agents (8 of each type)
• Timeout: 10 minutes
• Best for: Deep-dive analysis, exhaustive reports

DELIBERATION MODE (Multi-Round Debate)

Agents critique each other's answers and refine through peer review.

Trigger Patterns:

• "Deliberate on X"
• "Have the consultants debate X"
• "What do the models think about X after discussing?"
• "Peer review: X"

How It Works:

Round 1: Initial Perspectives (all 5 agents)

• Each agent provides initial analysis independently
• No agent sees others' responses yet

Round 2: Critique & Challenge (all 5 agents)

• Share all Round 1 responses with all agents
• Each agent reviews others' answers
• Point out errors, gaps, strong points
• Challenge assumptions
• Add missing information

Round 3: Refinement (all 5 agents)

• Share all critiques
• Agents revise their positions based on feedback
• Acknowledge valid points from others
• Defend or modify their stance
• Identify emerging consensus

Round 4: Final Synthesis (main session)

• Analyze convergence vs divergence
• Highlight consensus points (all/most agents agree)
• Present unresolved disagreements with reasoning from each side
• Rate confidence based on agent agreement

Timeout: 5 minutes total (agents work in rounds)

Best for:

• Critical decisions needing peer review
• Complex problems where single perspective is risky
• Catching errors through multiple reviews
• Finding consensus among experts
• Identifying trade-offs through debate

Example:

User: "Deliberate on: Should we use REST or GraphQL for our API?"

Round 1 (Initial):
- Codex: "GraphQL for flexible querying"
- Gemini: "REST for simplicity"
- Perplexity: [web research on adoption trends]

Round 2 (Critique):
- Codex: "Gemini's simplicity claim ignores client complexity - REST needs many endpoints"
- Gemini: "Codex didn't mention GraphQL's caching challenges"
- Perplexity: "Both missing recent data: GraphQL adoption is 45% in 2025"

Round 3 (Refinement):
- Codex: "Valid point on caching. Recommend REST for simple CRUD, GraphQL for complex reads"
- Gemini: "Agree with Codex's nuanced position"
- Consensus: Both viable, choose based on read complexity

Synthesis:
- CONSENSUS: Use REST for simple APIs, GraphQL for complex data fetching
- AGREEMENT: Both have trade-offs, no universal answer
- DISAGREEMENT: None (all converged)

Available Agents

• perplexity-researcher: Web search with Perplexity Sonar models
• claude-researcher: Web search with Claude WebSearch
• gemini-researcher: Analysis with Google Gemini
• codex-researcher: Deep analysis with GPT-5 Codex
• grok-researcher: Analysis with xAI Grok

How Consultation Mode Works

1. Detect consultation request from trigger patterns
2. Identify target model (gemini, codex, grok, perplexity, claude)
3. Launch single agent of that type
4. Return focused analysis from that one expert

Speed: ~10-30 seconds per consultation

How Research Mode Works

1. Query Decomposition: Break into 5-40 sub-questions
2. Parallel Launch: All agents in SINGLE message
3. Collection: Wait for timeout (2/3/10 minutes)
4. Synthesis: Integrate findings with confidence ratings

Speed:

• Quick: ~30-60 seconds
• Standard: ~30-90 seconds
• Extensive: ~1-3 minutes

Agent Capabilities

Web Search Agents:

• perplexity: Citations, Sonar models, current data
• claude: Built-in WebSearch, detailed analysis

LLM Analysis Agents:

• codex: GPT-5 with high reasoning, technical deep-dives
• gemini: Multi-perspective synthesis
• grok: Alternative LLM perspective

Best Practices

For Consultation:

• Use when you want ONE expert opinion
• Good for second opinions, alternative perspectives
• Faster and cheaper than research mode

For Research:

• Use when you need comprehensive coverage
• Multiple perspectives reveal blind spots
• Higher confidence through corroboration

Agent Selection:

• Codex/GPT-5: Complex technical problems, deep reasoning
• Gemini: Creative solutions, multi-angle analysis
• Grok: Alternative perspective, different training data
• Perplexity: Current web information, citations needed
• Claude: Web research, detailed synthesis