Agent Prompts

Agent Prompt Library

Canonical prompts live in src/Quickfire.Blazor/Domain/Agents/Utilities/PromptTemplates.cs. Keep this page in sync when prompts change so AI work stays deterministic.

Intent Classification

Used by OpenAIAgent.ClassifyIntentAsync before every request.

Classify this user input into one of four categories for an insurance business application:
1. AgentAction ...
2. DatabaseQuery ...
3. Navigation ...
4. GeneralAI ...
Input: "{{input}}"
Respond with ONLY raw JSON:
{
  "intent": "AgentAction|DatabaseQuery|Navigation|GeneralAI",
  "confidence": 0.0-1.0,
  "reasoning": "Brief explanation"
}

Tips

• Keep examples up to date with actual workflows (loss runs, DocuSign, SmartPaste, etc.)
• Avoid markdown/code fences so Semantic Kernel can parse JSON without trimming

SQL Generation

Executed inside HandleDatabaseQueryAsync to craft read-only SQL.

Generate a SQL query for this question: "{{input}}"
Database Schema:
{{schema}}
Requirements:
1. Use proper JOINs...
...
13. For phone number queries ...
Respond with ONLY raw JSON:
{
  "sql": "SELECT ...",
  "explanation": "What this query does",
  "success": true
}

House rules

• Keep business rules (client name vs ID, carrier aliasing, email joins) accurate with the real schema.
• Add new constraints when we add tables to keep SQL safe and performant.

Parameter Extraction

Used by Task Agents and SmartActions.

Extract parameters from this business action request: "{{input}}"
Extract these if present:
- client_name
- action_type (loss_run, certificate, proposal, payment_link, ...)
- policy_type
- time_period
- urgency
Respond with ONLY raw JSON (no markdown).

Extend the bullet list as agents learn new parameters (e.g., carrier_name, policy_number).

How to add/update prompts

1. Edit PromptTemplates.cs and add/adjust your template text. Keep placeholders wrapped in {{ }}.
2. Reference the new template in code via PromptTemplates.GetPrompt("TemplateName", new() { { "input", text } }).
3. Update this wiki page with the new template and guidance.
4. Store long-form prompt files (multi-hundred lines) under wwwroot/prompts and load them manually if they do not fit neatly inside PromptTemplates.

Testing

• Use /agents/enhanced/chat?debug=true to watch the actual prompt text, parsed response, and intent classification.
• For SQL prompts, run the question through the /api/debug/database-query endpoint and inspect server logs.
• Keep sample conversations in /ref/ToDo-AI.md so we can regression-test prompts before releasing.

Keeping prompts centralized here prevents drift between Cursor snippets, wiki docs, and the actual source.