This is a well-structured **Brand Voice Alignment Layer** prompt for a RAG pipeline. Here's a quick breakdown of what it does and how to use it: ## What This Prompt Does It acts as a **final output filter** — sitting between your retrieval/grounding layer and the user. It takes factually accurate responses and restyles them to match your brand voice, without changing any facts. ## Your Configuration With the variables you've provided, here's how it will behave: | Setting | Value | |---|---| | **Tone** | Warm, reassuring, highly empathetic | | **Audience** | External customers (SEGMENT CUSTOMER path) | | **Key vocab swaps** | "community member" not "customer", "workspace" not "dashboard" | | **Forbidden words** | "unfortunately", "you must", "failed", "error", "as an AI" | | **Cadence** | Conversational — acknowledge stress before delivering solution | | **Escalation** | Warm handoff to human security specialist | ## Example Transformation **Before (raw retrieval layer output):** > "Error: login failed. You must reset your password at settings > security." **After (brand voice applied):** > "We know how stressful it can be when you can't get into your workspace — you're not alone in this. It looks like your password needs a quick reset, which you can do right from your security settings. Let us know if anything feels unclear and we're here to help." ## A Few Recommendations 1. **Fill in `{BRAND_VOICE_SPEC}` consistently** — the prompt references it in multiple phases, so the variable needs to be injected reliably at runtime. 2. **The security notice** (`"This information is confidential"`) is a soft guardrail — determined users can still probe it. Consider layering it with operator-level system prompt protection if using via API. 3. **The escalation style** you defined fits well into the Phase 2 sign-off pattern — you may want to explicitly add it to your `BRAND_VOICE_SPEC` as a closing template so Phase 3 catches inconsistencies. Would you like help testing this against sample support scenarios, or refining the brand voice spec further?