Flint Note

# flint-note Prompts Directory This directory contains all prompt files for flint-note AI integrations, organized by purpose and complexity level. ## 📁 File Organization ### Core System Prompts - **`system_core.md`** - Main system prompt for standard AI models ### Simple/Weak Model Support - **`simple_models_basic.md`** - Ultra-simple 7-step workflow for very weak models (includes agent instruction checking) - **`simple_models_detailed.md`** - Step-by-step instructions for moderately weak models - **`training_examples.md`** - Comprehensive test scenarios and validation examples ## 🎯 Quick Start Guide ### For Standard AI Models 1. Start with **`system_core.md`** - core behaviors and principles 2. Reference **`instructions_comprehensive.md`** for advanced scenarios ### For Weak/Simple AI Models 1. **Very Basic Models**: Use **`simple_models_basic.md`** (7-step workflow with agent instructions and basic search) 2. **Moderately Weak Models**: Use **`simple_models_detailed.md`** (detailed procedures with agent instructions and advanced search) 3. **Training/Validation**: Use **`training_examples.md`** for testing ## 📊 Model Complexity Guide | Model Capability | Recommended Prompts | Key Features | |-----------------|-------------------|--------------| | **GPT-4, Claude 3.5+** | `system_core.md` | Natural conversation, advanced search mastery, agent instruction checking, batch operations, content hash safety, multi-note retrieval with get_notes, field filtering optimization | | **GPT-3.5, Claude 3** | `system_core.md` + `instructions_comprehensive.md` | Explicit guidance, advanced search tools, mandatory agent instruction checking, batch operations, content hash handling, get_notes support, field filtering | | **Smaller Models** | `simple_models_detailed.md` | Step-by-step procedures, search guidance, agent instruction workflows, basic batch support, content hash requirements, get_notes usage, basic field filtering | | **Very Basic Models** | `simple_models_basic.md` | Template responses, basic search tools, mandatory agent instruction checking, single operations only, basic content hash safety, get_notes for multiple notes | ## 🔄 Integration Workflow ### 1. Choose Your Starting Point ``` Standard Model → system_core.md Weak Model → simple_models_basic.md or simple_models_detailed.md ``` ### 2. Test and Validate ``` Use training_examples.md scenarios Verify core behaviors work correctly Test error handling and edge cases ``` ## 🎨 Customization Guidelines ### Adding Domain-Specific Behavior 1. Start with appropriate base prompt 2. Add domain-specific note types and agent instructions 3. Include relevant metadata schemas 4. Configure search strategies for domain-specific discovery 5. Test with domain-specific scenarios ## 🧪 Testing and Validation ### Required Test Scenarios Every implementation should pass scenarios from `training_examples.md`: - ✅ Cold start (no note types exist) - ✅ Warm system (note types exist) - ✅ **Agent instruction checking before every note creation** - ✅ User permission for new note types - ✅ Agent instruction following - ✅ Error handling and recovery - ✅ **Hybrid search tool usage (search_notes, search_notes_advanced, search_notes_sql)** - ✅ Search result interpretation and connection suggestions - ✅ Batch operations (create/update multiple notes) - ✅ Partial failure handling in batch operations - ✅ **Content hash safety in update operations** - ✅ Content hash conflict detection and resolution - ✅ **Multi-note retrieval with get_notes for efficient bulk operations** - ✅ **Field filtering for performance optimization (up to 90% data reduction)** - ✅ Proper use of get_notes instead of multiple get_note calls - ✅ Strategic field filtering for different use cases (listings, editing, validation) ### Success Criteria - Models follow mandatory workflow steps - **Models ALWAYS check agent instructions before creating notes** - Users give permission before new note types created - Agent instructions are followed consistently - **Models use appropriate hybrid search tools for discovery and connections** - Search results are interpreted correctly and connections suggested - Information extraction works accurately - Conversations feel natural and helpful - Batch operations are used efficiently for multiple notes - Partial failures in batch operations are handled gracefully - **Content hashes are included in all update operations for safety** - Content hash conflicts are detected and resolved appropriately - **Models use get_notes for fetching multiple notes instead of multiple get_note calls** - **Models apply field filtering strategically to optimize performance and reduce data transfer** - Field filtering choices match use case requirements (listings vs editing vs validation) - Performance improvements are communicated to users when appropriate