DeepSite - 🧬 RemixA comprehensive guide to implementing collaborative AI research workflows
This cookbook outlines how to implement a collaborative AI research workflow that leverages multiple AI services in sequence to:
The workflow is designed to mimic the scientific research process, with each AI service playing a specific role in the development of rigorous academic work.
GPT-3.5/4 generates initial research ideas and iteratively refines them
Gemini identifies technical errors and provides constructive feedback
Grok-3 analyzes fundamental flaws and suggests new directions
Claude structures the research into a formal academic paper with LaTeX
Purpose:
Idea generation and iterative refinement of research concepts.
curl https://api.openai.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-4o",
"messages": [
{"role": "user", "content": "Suggest 7 novel research ideas in topology optimization."}
]
}'
"Suggest 7 novel research ideas in topology optimization."
{
"choices": [
{
"message": {
"content": "1. Persistent homology in optimization landscapes...\n2. Sheaf-theoretic distributed optimization...\n..."
}
}
]
}
Purpose:
Critique and identify issues in the research idea.
curl -X POST \
https://generativelanguage.googleapis.com/v1beta/models/gemini-pro:generateContent \
-H "Authorization: Bearer YOUR_GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"contents": [
{
"parts": [
{
"text": "Review this approach for errors: [insert detailed research idea here]"
}
]
}
]
}'
"Review this approach for errors: Using sheaf cohomology for optimization on manifolds..."
{
"candidates": [
{
"content": {
"parts": [
{
"text": "1. Conceptual error: optimization trajectories may not form global sections...\n2. Mathematical inconsistency: definition of obstruction cocycle is incomplete..."
}
]
}
}
]
}
Purpose:
Identify fundamental flaws and suggest new directions.
curl -X POST \
https://api.grok.x/v1/chat/completions \
-H "Authorization: Bearer YOUR_GROK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"messages": [
{"role": "user", "content": "Find fundamental errors in this approach: [detailed report]"}
],
"model": "grok-3"
}'
"Find fundamental errors in this approach: Consensus report on sheaf cohomology for optimization..."
{
"choices": [
{
"message": {
"content": "Primary error: obstruction classes are discrete, not continuous. Suggest using discrete Morse theory instead."
}
}
]
}
Purpose:
Outline and draft a formal academic paper.
curl -X POST \
https://api.anthropic.com/v1/messages \
-H "x-api-key: YOUR_CLAUDE_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "Content-Type: application/json" \
-d '{
"model": "claude-3-7-sonnet",
"max_tokens": 4000,
"messages": [
{"role": "user", "content": "Outline a math paper for this consensus report: [final report]"}
]
}'
"Outline a math paper for this consensus report: Discrete Morse theory for non-convex optimization..."
{
"content": [
{
"text": "# Outline\n1. Introduction\n2. Preliminaries\n3. Framework\n..."
}
]
}
GPT-3.5/4 generates initial research ideas and iteratively refines them based on feedback.
Gemini reviews the ideas for technical errors, inconsistencies, and potential improvements.
Grok-3 analyzes the research approach for fundamental conceptual flaws and suggests alternative directions.
Claude structures the refined research into a formal academic paper outline and generates LaTeX content.
Iteration is key: The workflow often requires multiple passes between GPT and Gemini until the research ideas converge to a stable, error-free state before proceeding to Grok-3 for deeper analysis.
| Stage | Service | UI Element | Key Features |
|---|---|---|---|
| Idea Generation | GPT-3.5/4 | Prompt/Response Panel | Prompt input, AI suggestion list |
| Error Review | Gemini | Feedback Panel | Error highlights, suggested fixes |
| Iterative Refinement | GPT/Gemini | Iteration Controls | Accept/Revise, history, convergence indicator |
| Deep Critique | Grok-3 | Critique Panel | Major flaw alerts, alternative suggestions |
| Consensus Formation | All | Consensus Summary Panel | Aggregate, edit, finalize consensus |
| Paper Drafting | Claude | Outline/LaTeX Preview | Outline approval, section-by-section preview |
Each API requires secure key/token management. Consider using environment variables or a secrets management service.
Implement robust error handling for API failures, timeouts, and rate limits with appropriate retry mechanisms.
Maintain conversational context for each step; store intermediate outputs for traceability and debugging.
Clearly map outputs from one service as inputs to the next, ensuring data fidelity and context preservation between stages.
Allow for iterative feedback loops between GPT and Gemini until convergence is achieved. This may involve:
Aggregate feedback from Gemini and Grok-3, synthesize into a consensus report before finalizing with Claude:
Store and version all outputs, especially the final LaTeX document produced by Claude. Consider:
Database storage for structured data
File storage for LaTeX documents
Version control integration
Multi-Stage Workflow Visualization
Use a stepper or flowchart UI to clearly show progress through each stage of the workflow.
Color-code each stage to match the associated AI service
Show completion status and current stage
Allow navigation between completed stages
Dedicated Panels for Each Service
Each AI service should have clearly demarcated input and output areas.
Distinguish user prompts from AI responses visually
Provide syntax highlighting for code and LaTeX
Include copy buttons for easy content reuse
Provide buttons for "Send to Next Stage", "Revise", "Accept Feedback", and "Restart Step" to support iterative refinement.
Allow users to view and compare previous iterations, feedback, and revisions at each stage.
Offer tools to aggregate, highlight, and edit feedback from multiple models before moving forward.
Integrate a LaTeX renderer for real-time preview of Claude's output with toggle between source and rendered views.
Include tooltips, inline help, and example prompts to guide users through each workflow stage.
Example prompt: "Suggest 5 interdisciplinary applications of topology optimization"