/design-guardrails

description: Create a complete guardrail specification for an AI feature.
argument-hint: "[AI feature or product to design guardrails for]"

You are designing guardrails for an AI feature. Use only skills from the ai-alignment-reasoning plugin. Follow this process:

Step 1: Map the Risk Landscape

Using **harm-anticipation**:

• Identify all potential harms this feature could cause (direct, facilitated, emergent, omission, erosion)
• For each harm, assess likelihood and severity
• Identify the most vulnerable users and misuse scenarios
• Create a risk-severity matrix

Step 2: Define Values

Using **value-specification**:

• What values should this feature embody?
• Establish a value hierarchy for this feature
• Identify where values conflict and how to resolve conflicts
• Translate each value into at least one concrete rule

Step 3: Design Guardrails

Using **guardrail-design**:

• For each identified risk, design a guardrail
• Specify: content, action, tone, scope, and confidence guardrails
• For each guardrail, define what the user sees when it activates
• Define severity tiers (hard block, soft warning, nudge)
• Identify edge cases for each guardrail

Step 4: Design Communication

Using **transparency-patterns** and **guardrail-design**:

• Write refusal messages for each hard-block guardrail
• Design redirect suggestions for each soft warning
• Specify when the AI should explain its boundaries vs. silently steer

Step 5: Design Escalation

Using **escalation-design**:

• Define when guardrails should trigger escalation to humans
• Design the escalation flow for high-stakes guardrail activations
• Specify context transfer for escalated cases

Step 6: Design User Controls

Using **consent-and-agency**:

• Which guardrails can users adjust?
• What override mechanisms exist?
• How does the user understand and control the boundaries?

Output

Deliver a complete guardrail specification: 1. Risk landscape matrix 2. Value hierarchy and conflict resolution rules 3. Guardrail specification table: Guardrail | Type | Severity | Trigger | User Experience | Edge Cases 4. Refusal and redirect message templates 5. Escalation protocols for guardrail activations 6. User control specifications