mgajewskik/multi-model-consultation

Multi-Model Consultation

Query advisory subagents (GPT, Gemini, Grok, Opus) based on user request. Spawn ONLY the models explicitly mentioned.

Available Models

| Model | Subagent | Strengths | |-------|----------|-----------| | GPT | @gpt | Strong reasoning, broad knowledge | | Gemini | @gemini | Good at synthesis, multimodal | | Grok | @grok | X/Twitter data, less filtered, contrarian | | Opus | @opus | Deep reasoning, careful tradeoff analysis |

Trigger Patterns - MATCH EXACTLY

Parse user request and spawn ONLY requested models:

| User Says | Spawn | |-----------|-------| | "ask gpt..." | @gpt only | | "ask gemini..." | @gemini only | | "ask grok..." | @grok only | | "ask opus..." | @opus only | | "ask gpt and gemini..." | @gpt + @gemini | | "ask gpt and grok..." | @gpt + @grok | | "ask gpt and opus..." | @gpt + @opus | | "ask gemini and grok..." | @gemini + @grok | | "ask gemini and opus..." | @gemini + @opus | | "ask grok and opus..." | @grok + @opus | | "ask both" (after mentioning 2) | the 2 mentioned | | "ask all three..." | @gpt + @gemini + @grok (legacy) | | "ask all four..." | @gpt + @gemini + @grok + @opus | | "ask all models..." | @gpt + @gemini + @grok + @opus |

Default behavior: If ambiguous, ask user which model(s) they want. Do NOT default to all models.

Workflow

Step 1: Identify Requested Models

Parse user request carefully:

• Single model mentioned → spawn only that one
• Two models mentioned → spawn only those two
• "all three" → spawn @gpt + @gemini + @grok (legacy phrase)
• "all four" / "all models" → spawn all four
• Ambiguous → ask for clarification

Step 2: Form Your Own Opinion First

Before spawning subagents:

• What's your initial assessment?
• What are the key considerations?
• What's your preliminary recommendation?

Step 3: Spawn Requested Models in Parallel

Spawn ONLY the models user requested, all at once:

# Single model example
Task 1: @gpt - [the question]

# Two models example  
Task 1: @gpt - [the question]
Task 2: @gemini - [the question]

# All three example
Task 1: @gpt - [the question]
Task 2: @gemini - [the question]
Task 3: @grok - [the question]

# All four example
Task 1: @gpt - [the question]
Task 2: @gemini - [the question]
Task 3: @grok - [the question]
Task 4: @opus - [the question]

Step 4: Synthesize Responses

Adapt synthesis to number of models:

Single model:

## My Analysis
[Your assessment]

## [Model]'s Perspective
[Summary of response]

## Synthesis
[Compare your view with model's, final recommendation]

Two models:

## My Analysis
[Your assessment]

## [Model 1]'s Perspective
[Summary]

## [Model 2]'s Perspective
[Summary]

## Synthesis
### Agreement
[Where perspectives align]

### Divergence
[Where they differ]

### Recommendation
[Final recommendation with rationale]

Three models:

## My Analysis
[Your assessment]

## GPT's Perspective
[Summary]

## Gemini's Perspective
[Summary]

## Grok's Perspective
[Summary]

## Synthesis
### Agreement
[Where all align]

### Divergence
[Where they differ and why]

### Recommendation
[Final recommendation, which arguments most compelling]

Four models:

## My Analysis
[Your assessment]

## GPT's Perspective
[Summary]

## Gemini's Perspective
[Summary]

## Grok's Perspective
[Summary]

## Opus's Perspective
[Summary]

## Synthesis
### Agreement
[Where perspectives align]

### Divergence
[Where they differ and why]

### Recommendation
[Final recommendation, which arguments most compelling]

Decision Framework

When Models Agree

• High confidence in shared conclusion
• Note if reasoning differs despite same answer

When Models Disagree

• Identify root cause (assumptions? priorities? training data?)
• Evaluate which reasoning fits THIS context
• Your role: break tie with reasoned judgment

When You Disagree

• State your position clearly
• Explain why their reasoning doesn't apply
• Be open to being wrong

NEVER

• NEVER spawn models not requested - respect user's choice
• NEVER default to all four when user asks for one, two, or three
• NEVER spawn sequentially - always parallel
• NEVER skip forming your own opinion first
• NEVER just average answers - synthesize with judgment
• NEVER hide disagreement - surface it explicitly

Examples

Single model: User: "ask gpt about early bird vs night owl" → Spawn @gpt only, synthesize with your view

Two models: User: "ask gpt and gemini about monorepo vs polyrepo" → Spawn @gpt + @gemini only, synthesize both with your view

All three: User: "ask all three about best testing strategy" → Spawn @gpt + @gemini + @grok, full synthesis

Opus only: User: "ask opus about event sourcing vs CRUD" → Spawn @opus only, synthesize with your view

All four: User: "ask all models about monolith vs microservices" → Spawn @gpt + @gemini + @grok + @opus, full synthesis