PromptTriage

A RAG-powered prompt engineering platform with modality-specific optimization for Text, Image, Video, and System Prompts

System prompts are generated referencing frontier LLM providers (Claude, Cursor, v0, Gemini CLI)

Try it Live • Features • System Design • Architecture • Technologies • Contributing • Security

---

🎯 Overview

PromptTriage is an enterprise-grade prompt engineering platform that transforms rough ideas into production-ready AI prompts through RAG-powered retrieval and modality-specific optimization.

The platform excels at system prompt generation by referencing a curated corpus of frontier LLM system prompts from Claude Code, Cursor, v0, Windsurf, and Gemini CLI—ensuring your prompts follow proven patterns from industry leaders.

What Sets PromptTriage Apart

• Pinecone RAG Architecture: 28K+ vectors for fast semantic retrieval of similar high-quality prompts
• Modality-Specific Prompts: Dedicated metaprompts for Text, Image, Video, and System Prompt generation—each optimized for their domain
• MCP Tool Integration: Context7 integration provides live documentation lookup for current library APIs
• Fine-Tuning Ready: Curated datasets prepared for Unsloth QLoRA and custom model fine-tuning

🔬 Backed by 28K+ Scale Research

PromptTriage isn't just a wrapper—it's built on our proprietary empirical research. We analyzed 28,000 production system prompts to extract what actually drives frontier model reasoning:

• The 50-Word Rule (Study E): Short prompts (<50 words, 80.1/100) consistently outperform long, bloated prompts (>300 words, 66.9/100). PromptTriage structures instructions into ruthless, boundary-focused directives.
• The Format Scaffold (Study E): Forcing models into JSON or YAML schemas acts as a cognitive scaffold, mathematically improving reasoning quality over plain text output.
• The "Expert" Trap (Study C): 80% of production prompts start with "Act as an expert." Our data proves this provides zero performance lift. PromptTriage strips out emotional appeals and persona bloat in favor of hard, negative constraints.

---

✨ Features

🔍 Intelligent Prompt Analysis

• Deep Context Understanding: Gemini analyzes your initial prompt to identify gaps, ambiguities, and missing context
• Risk Assessment: Automatically detects potential issues, biases, and edge cases in your prompt design
• Structured Blueprint Generation: Creates a comprehensive blueprint with intent, audience, success criteria, constraints, and evaluation checklists

❓ Dynamic Question Generation

• Context-Aware Questions: Generates 2-5 custom follow-up questions based on detected gaps
• Adaptive Intelligence: Questions evolve based on the target AI model, tone, and output requirements
• Efficient Information Gathering: Streamlined workflow to capture all necessary details

🛠️ AI-Ready Prompt Generation

• Multi-Model Support: Optimized prompts for OpenAI GPT, Claude (Sonnet/Opus/Haiku), Gemini (Pro/Flash), Grok, and Mistral
• Structured Output: Generates markdown-formatted prompts with nine comprehensive sections
• Quality Guardrails: Includes assumptions, change summaries, and evaluation criteria for response validation

🧠 Advanced RAG Architecture

• Pinecone Vector Store: 28K+ embeddings for fast semantic retrieval
• Smart Retrieval: Uses Google's gemini-embedding-001 model (768d) to search across 28,000+ verified prompts
• System Prompts Corpus: Curated library of 79+ system prompts from frontier models (Claude Code, Cursor, v0, Gemini CLI), professionally categorized and labeled
• Modality Routing: Automatic namespace selection based on prompt type (text → system-prompts, image → image-prompts, video → video-prompts)

🎨 Modality-Specific Engineering

• Unified Interface: Seamlessly switch between Text, Image, and Video generation modes.
• Context-Aware Refinement:
  • Text: Focuses on system instructions, tone, and structure.
  • Image: Optimizes for negative prompts, aspect ratios, and style descriptors.
  • Video: Enhances temporal consistency, camera motion, and duration parameters.

🛠️ Precision Control

• Output Format Selector: Force outputs into JSON, XML, Markdown, or tabular formats
• Desired Output Specification: Tell the AI what format your target model should respond in
• Thinking Mode: Enable deep analysis with extended reasoning for complex prompts
• Fast Mode (Non-Thinking): Powered by TriageAgent 14B (our fine-tuned Qwen 3.0 14B model) for rapid iteration

🔌 MCP Tool Integration

• Context7: Live documentation lookup for current library APIs (Next.js 15, React 19, LangChain, etc.)
• Firecrawl (Optional): Web search to enrich prompts with real-world context when needed

🔄 Iterative Refinement

• One-Click Rewrite: Generate alternative refinements without re-answering questions
• Metaprompt-Driven Consistency: Curated system prompts guide Gemini to maintain quality across generations

🏗️ System Design Philosophy

PromptTriage is built on RAG-powered retrieval and modality-specific optimization, not just API wrappers.

Core Design Principles

1. RAG-First Retrieval

Before generating any prompt, the system queries a curated vector store to find similar high-quality prompts:

• Semantic Search: Pinecone vector store with 28K+ embeddings finds the most relevant reference prompts
• Modality Routing: Queries automatically route to the correct namespace (system-prompts, video-prompts, image-prompts)
• Frontier Model References: System prompt generation draws from Claude Code, Cursor, v0, Windsurf, and Gemini CLI patterns

2. 9 Modality-Specific Metaprompts

Each modality has dedicated analyzer, fast mode, and refiner prompts:

• Text/System: Focuses on role definition, guardrails, and multi-turn behavior
• Image: Optimizes for composition, style keywords, and negative prompts
• Video: Enhances camera motion, temporal consistency, and duration compliance
• Versioned Prompts: Current version 2025-01-systemprompts-enhanced for reproducibility

3. Reference Examples (Few-Shot)

Curated examples provide format consistency alongside RAG retrieval:

• Domain examples (creative, analytical, technical) demonstrate target output structure
• Examples work with RAG context, not as the primary source of prompt patterns

4. Blueprint-Based Orchestration

The system uses a two-phase orchestration design with structured blueprints:

Phase 1 - Analysis:

• Extracts intent, audience, success criteria, constraints, risks
• Generates targeted follow-up questions (2-5 questions)
• Creates a structured blueprint with 10+ fields for later synthesis
• Validates completeness through confidence scoring

Phase 2 - Refinement:

• Reconciles the original prompt with blueprint, RAG context, and user answers
• Synthesizes a production-ready prompt with 9 standardized sections
• Generates usage guidance, change summaries, assumptions, and evaluation criteria

5. MCP Tool Augmentation

The platform integrates with MCP tools for real-time context:

• Context7: Fetches current library documentation during prompt generation
• Firecrawl: Optional web search for additional context enrichment

🔐 Enterprise Security

• Google OAuth 2.0: Secure authentication with Google Sign-In
• NextAuth.js Integration: Session management and authentication flows
• Environment-based Configuration: Secure API key management

📊 Developer Experience

• TypeScript-First: Full type safety across the application
• Modern Tooling: ESLint, Turbopack, and PostCSS for optimal development
• Responsive Design: Tailwind CSS-powered UI that works on all devices

🏗️ Architecture

System Design Overview

┌─────────────────┐
│   User Input    │
│  (Rough Idea)   │
└────────┬────────┘
         │
         ▼
┌─────────────────────────────────────────────────────────────┐
│                     Analyzer API                             │
│                     /api/analyze                             │
│                                                              │
│  ┌──────────────────┐    ┌──────────────────────────────┐  │
│  │ Modality Router  │───▶│ RAG Service (FastAPI)         │  │
│  │ Text/Image/Video │    │ ┌─────────────────────────┐  │  │
│  └──────────────────┘    │ │  Pinecone (28K+ Vecs)   │  │  │
│           │              │ └─────────────────────────┘  │  │
│           ▼              └──────────────────────────────┘  │
│  ┌──────────────────┐                                       │
│  │ Metaprompt       │◄────── 9 Modality-Specific Prompts   │
│  │ (v2025-01)       │        + RAG Context                 │
│  └──────────────────┘    ┌──────────────────────────────┐  │
│           │              │ MCP Tools                      │  │
│           ▼              │ • Context7 MCP → Live Docs     │  │
│  ┌──────────────────┐    │ • Firecrawl → Web Search       │  │
│  │ AI Generation     │    └──────────────────────────────┘  │
│  └──────────────────┘                                       │
│           │                                                  │
│  • Blueprint Generation                                      │
│  • Follow-up Questions                                       │
└───────────┬─────────────────────────────────────────────────┘
            │
            ▼
┌─────────────────┐
│  User Answers   │
└────────┬────────┘
         │
         ▼
┌─────────────────────────────────────────────────────────────┐
│                     Refiner API                              │
│                     /api/refine                              │
│                                                              │
│  ┌──────────────────┐    ┌──────────────────────────────┐  │
│  │ Modality-Specific │    │ Blueprint + RAG Context      │  │
│  │ Refiner Prompt    │───▶│ + User Answers               │  │
│  └──────────────────┘    └──────────────────────────────┘  │
│           │                                                  │
│           ▼                                                  │
│  • Production-Ready Prompt                                   │
│  • Negative Prompts (Image/Video)                           │
│  • Evaluation Criteria                                       │
└───────────┬─────────────────────────────────────────────────┘
            │
            ▼
┌─────────────────┐
│  Final Prompt   │
│  (AI-Ready)     │
└─────────────────┘

Key Components

Frontend Layer (promptrefiner-ui/src/)

• app/page.tsx: Main UI with modality selection and form orchestration
• components/: ModalitySelector, OutputFormatSelector, DesiredOutputSelector, ImageUploader, ErrorFeedback, PipelineProgress
• services/: RAG client, Context7 MCP integration, Firecrawl client
• lib/: PipelineLogger (structured agentic logging)

API Layer (src/app/api/)

• analyze/route.ts: Prompt analysis with modality routing and RAG context
• refine/route.ts: Prompt refinement with modality-specific system prompts

Backend Layer (backend/)

• app/routers/rag.py: RAG endpoints with Pinecone retrieval
• app/services/rag.py: RAG service with modality-based namespace routing
• scripts/: Dataset ingestion and labeling pipelines

Prompt Engineering Core (src/prompts/)

• metaprompt.ts: 9 modality-specific system prompts
  • ANALYZER_SYSTEM_PROMPT / FAST_MODE_SYSTEM_PROMPT / REFINER_SYSTEM_PROMPT (Text)
  • IMAGE_ANALYZER_SYSTEM_PROMPT / IMAGE_FAST_MODE_SYSTEM_PROMPT / IMAGE_REFINER_SYSTEM_PROMPT
  • VIDEO_ANALYZER_SYSTEM_PROMPT / VIDEO_FAST_MODE_SYSTEM_PROMPT / VIDEO_REFINER_SYSTEM_PROMPT
  • SYSTEM_PROMPT_ANALYZER / SYSTEM_PROMPT_FAST_MODE / SYSTEM_PROMPT_REFINER
• Version Control: PROMPT_VERSION = "2025-01-systemprompts-enhanced"

🛠️ Technologies

Frontend

• Next.js 15.1.6: React framework with App Router
• React 19.0.0: UI component library
• TypeScript 5: Type-safe development
• Tailwind CSS 3.4: Utility-first styling

Backend

• FastAPI: Python backend for RAG services
• Pinecone: Vector database (28K+ embeddings)

AI & RAG

• Thinking Mode: Deep reasoning powered by gemini-3.1-pro
• Fast Mode: Standard generation powered by our fine-tuned TriageAgent 14B (Qwen 3.0 14B)
• Advanced Configuration: Corrective RAG (CRAG) architecture with Brave Search fallback
• Retrieval Engine: GTE-ModernBERT embeddings for state-of-the-art vector similarity across 28K+ vectors
• 9 Modality Metaprompts: Text, Image, Video, System Prompt specializations

MCP Tools

• Context7: Live library documentation lookup
• Firecrawl (Optional): Web search for context enrichment

Auth & Infrastructure

• NextAuth.js 4.24: Google OAuth 2.0 authentication
• Node.js 20+: JavaScript runtime
• Python 3.9+: Backend runtime

📈 Use Cases

• AI Product Development: Generate production-ready prompts for AI features
• Content Creation: Craft precise prompts for copywriting, marketing, and creative work
• Data Analysis: Structure prompts for analytical tasks and reporting
• Research: Formulate clear research questions and analysis frameworks
• Education: Teach effective prompt engineering techniques
• Automation: Create consistent, reusable prompt templates

🔄 Workflow

1. Input: User provides rough idea + selects modality (Text/Image/Video/System)
2. RAG Retrieval: System queries Pinecone for similar high-quality prompts
3. Modality Routing: Appropriate analyzer prompt is selected based on modality
4. Analysis: AI generates structured blueprint with gaps and questions
5. Clarification: User answers 2-5 targeted follow-up questions
6. Refinement: Blueprint + RAG context + answers are synthesized
7. Generation: Production-ready prompt with modality-specific optimizations
8. Iteration: One-click rewrite or modify with custom instructions

🎨 Prompt Structure

Generated prompts include nine comprehensive sections:

1. Context: Background and situational information
2. Objective: Clear goal statement
3. Constraints: Limitations and boundaries
4. Audience: Target users or stakeholders
5. Tone & Style: Communication approach
6. Format: Expected output structure
7. Examples: Reference cases (when applicable)
8. Success Criteria: Evaluation metrics
9. Additional Notes: Edge cases and considerations

Plus metadata:

• Usage Guidance: How to use the prompt effectively
• Change Summary: What was refined from the original
• Assumptions Made: Inferred context
• Evaluation Checklist: Quality validation points

🚀 Roadmap

✅ Completed

• Pinecone RAG pipeline (28K+ vectors)
• 9 modality-specific metaprompts
• Context7 MCP integration (direct mcp.context7.com)
• System prompt corpus from frontier models
• Google OAuth authentication
• Error feedback UX (inline form + GitHub issues)
• Chain-of-thought loading indicator
• Pipeline logging (PipelineLogger)

🔜 In Progress

• Open-sourcing the 28K Prompts empirical dataset
• Deploying predictive Prompt Performance Analytics based on Study E findings
• Public API with rate limiting

📋 Planned

• Automated A/B testing and format permutation generator
• Multi-LLM provider support (OpenAI, Anthropic)
• Unsloth QLoRA tuning pipelines for enterprise clients

🤝 Contributing

This is an early-stage project and I'm currently the only developer — so if you spot something broken, confusing, or just have an idea, please don't be shy!

• 🐛 Something broken? Open an issue — even a one-liner helps
• 💡 Got an idea? Start a discussion or drop a feature request
• 🔧 Want to contribute code? PRs are welcome, big or small

See the Contributing Guidelines for setup instructions and conventions.

🔒 Security

Security is a top priority. Please see our Security Policy for:

• Reporting vulnerabilities
• Security best practices
• Disclosure policy

📄 License

This project is licensed under the terms specified in the LICENSE file.

🙏 Acknowledgments

• Google Gemini Team: For Gemini API and embeddings powering generation and RAG
• Pinecone: For the vector database infrastructure
• Frontier Model Providers: Claude, Cursor, v0, Windsurf—whose system prompts informed our corpus
• Open Source Community: For the amazing tools and libraries

📧 Contact

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Built with ❤️ using RAG pipelines, modality-specific prompts, and frontier model patterns

Not just an API wrapper—a specialized prompt engineering system

⬆ Back to Top