Alert optimizer

Alert Optimizer

Overview

This skill takes error analysis data (ideally from the error-monitoring skill) and generates optimized alert rules, severity tiers, escalation policies, and on-call runbooks. It turns a noisy alerting setup into a structured incident response system.

Instructions

1. Understand Current State

Ask for or infer:

• Current monitoring platform (Sentry, Datadog, PagerDuty, etc.)
• Current alert volume and on-call team size
• Notification channels available (Slack, PagerDuty, email, SMS)
• Any existing severity definitions

2. Define Severity Tiers

Create a three-tier model (unless the user specifies otherwise):

Tier	Criteria	Response Time	Channel
P1 - Critical	Revenue impact, data loss, security breach, >50% users affected	Immediate page	PagerDuty/SMS
P2 - Warning	Degraded experience, >5% users affected, error rate spike	1 hour	Slack channel
P3 - Info	Known issues, cosmetic errors, self-healing transients	Weekly review	Log only

3. Generate Alert Rules

For each error group, produce a platform-specific alert configuration:

• Sentry: JSON alert rule with conditions, filters, and actions
• Datadog: Monitor definition with query, thresholds, and notification targets
• PagerDuty: Event rules with severity mapping and escalation policy
• Generic: Webhook payload template with routing logic

4. Create Escalation Policies

Define who gets notified and when:

• P1: On-call engineer immediately → team lead after 10 min → engineering manager after 30 min
• P2: Post to team Slack channel → on-call acknowledges within 1 hour
• P3: Aggregated weekly digest

5. Generate Runbooks

For each P1 alert, create a runbook with:

• What: One-sentence description of the alert
• Why it matters: Business impact
• Diagnose: First 3 steps to investigate
• Fix: Common resolutions
• Escalate: When and to whom

Examples

Example 1: Sentry Alert Rules

Input: Error audit showing PaymentProcessingError as P1, RedisTimeout as P3.

Output:

{
  "rules": [
    {
      "name": "P1: Payment Processing Failure",
      "conditions": [
        {"id": "sentry.rules.conditions.event_frequency", "value": 1, "interval": "5m"}
      ],
      "filters": [
        {"id": "sentry.rules.filters.tagged_event", "key": "error_type", "value": "PaymentProcessingError"}
      ],
      "actions": [
        {"id": "sentry.integrations.pagerduty.notify_action", "service": "engineering-oncall"},
        {"id": "sentry.integrations.slack.notify_action", "channel": "#incidents", "tags": "environment,transaction"}
      ],
      "frequency": 5
    },
    {
      "name": "P3: Redis Connection Timeout (Log Only)",
      "conditions": [
        {"id": "sentry.rules.conditions.event_frequency", "value": 100, "interval": "1h"}
      ],
      "filters": [
        {"id": "sentry.rules.filters.tagged_event", "key": "error_type", "value": "TimeoutError"}
      ],
      "actions": [],
      "frequency": 1440
    }
  ]
}

Example 2: On-Call Runbook Entry

Input: P1 alert for database deadlocks.

Output:

## RUNBOOK: Database Deadlock Detected

**Alert**: DatabaseError: deadlock detected in order processing
**Severity**: P1 — Revenue impact, orders fail silently
**Response**: Immediate — acknowledge within 5 minutes

### Diagnose
1. Check current DB connections: `SELECT * FROM pg_stat_activity WHERE state = 'active' ORDER BY query_start;`
2. Identify locked queries: `SELECT * FROM pg_locks WHERE NOT granted;`
3. Check if deployment happened in last 30 min (new query pattern?)

### Common Fixes
- Kill the blocking query: `SELECT pg_terminate_backend(<pid>);`
- If recurring post-deploy: rollback last migration
- If load-related: enable connection pooling or add read replica

### Escalate
- If not resolved in 15 min → page team lead
- If data corruption suspected → page engineering manager + DB admin

Guidelines

• Always deduplicate alert rules — one root cause should trigger one alert, not five
• Set reasonable frequency caps: P1 alerts should re-fire every 5 minutes max, P3 should be daily at most
• Include "auto-resolve" rules where appropriate (e.g., error rate drops below threshold)
• Runbooks should be copy-pasteable — include actual commands, not pseudocode
• When the user has fewer than 3 people on-call, simplify escalation to two tiers
• Test configurations by asking the user to dry-run before applying