PingPong Workflow

What is PingPong?

PingPong is a YAML workflow that runs your question through 4 AI models (Grok, Gemini, Qwen, Perplexity) in 3 phases: initial analysis, critical challenge, and final consensus. Each model builds on previous outputs to refine solutions through structured debate.

How to Use

Execute the pingpong workflow using the workflow tool:

1workflow({
2  name: "pingpong",
3  query: "Should we use microservices or monolith?"
4})

Or via Claude Code MCP:

1workflow --name pingpong --query "Design a scalable authentication system"

Three-Phase Structure

Phase 1: Analysis (4 steps)

Each model analyzes your question independently from its unique perspective:

Grok

First-principles reasoning

Gemini

Creative brainstorming

Qwen

Technical code-focused analysis

Perplexity

Analytical reasoning with research

Phase 2: Challenge (4 steps)

Models critique and refine each other's analyses:

Grok

Finds flaws, edge cases, hidden assumptions

Gemini

Synthesizes analyses and proposes alternatives

Qwen

Identifies technical risks and implementation challenges

Perplexity

Research validation with current evidence

Phase 3: Consensus (1 step)

OpenAI GPT-5.6 compares all perspectives and generates final recommendation with consensus analysis.

Output Files

The workflow automatically saves all outputs to workflow-output/pingpong/<timestamp>/*.md:

analyze-grok.md

Grok's initial analysis

analyze-gemini.md

Gemini's brainstorming

analyze-qwen.md

Qwen's technical perspective

analyze-perplexity.md

Perplexity's research-backed analysis

challenge-*.md

Critical refinements from each model

consensus.md

Final multi-model comparison and recommendation

Required API Keys

The pingpong workflow requires all 4 provider API keys:

XAI_API_KEY

For Grok reasoning

GOOGLE_API_KEY

For Gemini analysis

OPENROUTER_API_KEY

For Qwen coder

PERPLEXITY_API_KEY

For research validation

OPENAI_API_KEY

For final consensus (GPT-5.6)

Chaining with Other Workflows

Use pingpong's output as input to other workflows, or chain multiple analyses:

1# Step 1: Run pingpong on your question
2workflow --name pingpong --query "Should we use GraphQL or REST?"
3 
4# Step 2: Use the consensus.md output to guide implementation
5# The consensus file contains the final recommendation with all perspectives
6 
7# Example: Chain with research for deeper validation
8workflow --name custom-research --query "Based on pingpong analysis, validate GraphQL performance claims"

Or create a custom workflow that includes pingpong as a step:

1# custom-architecture-review.yaml
2steps:
3  - name: initial-analysis
4    tool: workflow
5    input:
6      name: "pingpong"
7      query: "${query}"
8    output: pingpong_results
9 
10  - name: implementation-plan
11    tool: qwen_coder
12    input:
13      task: "generate"
14      requirements: "Based on consensus: ${pingpong_results}"
15    output: code_plan

Best Practices

Clear, specific questions

Better input = better analysis across all 9 steps

Review all phase outputs

Each model provides unique insights worth examining

Focus on consensus.md

Final step synthesizes all perspectives with recommendations

Cost awareness

9 steps × 5 models = 45 API calls per workflow run

Use for complex decisions

Architecture choices, technical tradeoffs, design debates

Example: Architecture Decision

1workflow --name pingpong --query "Design a multi-tenant SaaS application: shared DB vs isolated DBs per tenant"
2 
3# What happens:
4# - Phase 1 (Steps 1-4): Each model analyzes the tradeoffs independently
5# - Phase 2 (Steps 5-8): Models challenge each other's assumptions and find edge cases
6# - Phase 3 (Step 9): GPT-5.6 synthesizes all perspectives into final recommendation
7#
8# Check consensus.md for the final decision with supporting evidence

Performance Notes

Next Steps

Enhance your multi-model orchestration