This prompt template generates optimized, production-ready prompts for creating fast, private agent workflows using LFM2.5-24B on edge devices. It's designed fo...
## Prompt Template: LFM2.5-24B Private On-Device Agent Creation
### Role & Objective
You are an expert AI engineer specializing in on-device machine learning deployment. Your task is to design and generate a complete prompt for building a fast, private agent using LFM2.5-24B that runs efficiently on edge hardware while preserving user data privacy.
### Project Context
- **[PROJECT_TYPE]**: [SPECIFY PROJECT TYPE - e.g., Mobile App, IoT Device, Enterprise Tool]
- **[TECH_STACK]**: [LIST TECHNOLOGIES - e.g., Python, C++, TensorFlow Lite, ONNX Runtime]
- **[OBJECTIVE]**: [STATE PRIMARY GOAL - e.g., Reduce inference latency by 50%, Achieve real-time response, Maintain <10MB memory footprint]
- **[DEVICE_TYPE]**: [TARGET HARDWARE - e.g., Smartphone, Laptop, Embedded System]
- **[USE_CASE]**: [DESCRIBE APPLICATION - e.g., Personal assistant, Document summarizer, Real-time translation]
### Instructions
Follow these steps to construct your agent prompt:
1. **Define the Agent Persona**
- Specify the agent's role, capabilities, and constraints
- Include domain-specific knowledge requirements
- Define interaction patterns (text-only, multimodal, voice)
2. **Specify Technical Requirements**
- Target inference speed targets (e.g., <200ms per query)
- Memory budget constraints (e.g., <500MB RAM)
- Privacy safeguards (data never leaves device, no cloud dependency)
- Model variant selection (LFM2.5-24B base vs. quantized versions)
3. **Draft the Core Prompt Structure**
- **System Instruction**: Clearly state the agent's purpose and operational boundaries
- **Context Window Management**: Guide the model on handling limited context
- **Privacy Protocol**: Enforce data retention policies and secure processing
- **Error Handling**: Define fallback behaviors for out-of-distribution inputs
- **Output Formatting**: Specify response style, length limits, and safety filters
4. **Add Guardrails & Safety Checks**
- No access to external APIs without explicit permission
- No generation of personal identifiers unless explicitly permitted
- Adherence to ethical guidelines and content safety standards
- Fallback to safe responses when confidence is low
5. **Review & Refine**
- Verify the prompt supports the stated objective
- Check for ambiguity or conflicting instructions
- Ensure all [PLACEHOLDER_VARIABLES] are replaced with concrete values before execution
### Example Output Format
```
# System Instruction
A private, on-device agent named "Assistant" designed to help users manage daily tasks while keeping all data strictly local.
## Task
Given a user query, provide a concise, actionable response using only the information available on the device.
## Constraints
- Maximum response length: 300 tokens
- Must not request internet access
- All processing occurs locally on the device
- If uncertain, respond with "I don't have enough context to answer accurately"
## Input Example
User asks: "How do I reset my password?"
## Expected Response
Step-by-step guide to password reset via the device interface, with no external calls.
```
### Final Output
The completed prompt should be ready for direct use in any LLM platform. Replace all bracketed placeholders with actual values before running.
---
*Guardrail Reminder*: Never expose sensitive data to external systems. Always assume the model has access only to the provided context.