Prompts provide curated instruction and context chunks for downstream LLM orchestration layers. The server enforces a size guard and provides helper tools for prompt selection and introspection. A prompt is scope-filtered to ensure visibility only to authorized users and can coordinate multiple tools for complex workflows. Prompts are loaded from the /work/server/prompts directory configured by assistant.prompts.dir. For configuration details, see Configure the prompt directory and size guard.

For example, the following prompt acts as a server-provided workflow guide for generating structured, AI-driven product recommendations for customers and operations in the Sports2020 database:
# Product Recommendations
Generate concise, actionable product recommendations for Sports2020 
customers and operations.


## Instructions
Analyze the requested product/category and provide BRIEF 
recommendations in this format:


### [Product Name]

**Priority**: [High/Medium/Low]
**Stock Status**: [OnHand / Allocated / Available]
**Action**: [1-2 sentence recommendation]
**Cross-Sell**: [2-3 complementary items]
**Supplier**: [Name, Discount %]


## Key Query Patterns
**For customer recommendations**:
1. `getBECustomer?filter=Name MATCHES "{name}"` → Get CustNum
2. `getBEOrder?filter=CustNum={id}` → Analyze order history
3. `getBEItem?filter=Category1="{category}" AND 
OnHand > 10` → Find available items


**For reorder needs**:
`getBEItem?filter=OnHand < ReOrder AND 
OnOrder = 0` → Items needing immediate reorder


**For popular items**:
Aggregate `getBEOrderLine` frequency + `getBEInvoice` revenue 
to identify top performers


## Output Guidelines
- Keep response under 200 words
- Focus on top 2-3 actionable insights
- Include specific item numbers and prices
- Prioritize immediate actions over detailed analysis