Create Policies with AI

Describe a FinOps rule in plain English and SOFE generates a valid policy YAML automatically.

How It Works

You type: "Alert when S3 buckets don't have encryption enabled"
    ↓
AI generates valid SOFE policy YAML:
    ↓
apiVersion: sofe/v1
kind: Policy
metadata:
  name: s3-encryption-required
  description: "Detect S3 buckets without encryption enabled"
  severity: high
spec:
  scope:
    resource_types: ["aws.s3"]
  rule:
    metric: encryption_enabled
    operator: "=="
    threshold: 0
  message: "encryption_enabled = {value} (threshold: =={threshold})"
  recommendation: "Enable default encryption (SSE-S3 or SSE-KMS)"
    ↓
You review → Save → Policy is yours

Prerequisites

• AI Provider configured (see guide)
• Recommend a ≥30B model (Claude, Llama 70B, Gemma 31B) for accurate YAML generation

Step-by-Step

1. 1. Go to platform.sofe.dev/policies/create
2. 2. Type your policy description in plain English (min 10 characters)
3. 3. Click "🧠 Generate Policy" — wait 3-5 seconds
4. 4. Review the generated YAML (shown in green preview)
5. 5. Click "💾 Save Policy" to store it
6. 6. Find it in /catalog → Policies tab → "My Custom Policies"

Example Prompts

Available Metrics

Your policy can only use metrics that SOFE collectors actually measure. Current metrics:

If you request a policy for a metric that doesn't exist (e.g., "spot instance price"), the AI will respond with an error explaining the limitation.

Validation

Generated YAML is validated before showing it to you:

• Must contain apiVersion: sofe/v1, kind: Policy, spec:
• If invalid, AI retries automatically (up to 3 attempts)
• If still invalid after 3 attempts, you get an error message

Saved Policies

Saved policies appear in:

• /catalog → Policies tab → "My Custom Policies" section
• Stored in your Firestore account (private to you)
• Coming soon: Custom policies will be included in cloud evaluations automatically

Limitations

• Only available metrics can be used (see list above)
• Custom policies not yet included in scheduled evaluations (coming soon)
• Each generation counts as 1 AI call (rate limited)
• Weaker models (<30B) may generate invalid YAML — use recommended models