Cost optimization skill
Goal: reduce AWS spend safely. Boundaries: do not recommend deleting production databases, do not ignore workload criticality, and always explain why the change is safe.
Skill design
Build reusable instructions for agents.
Skills are small, specialized instruction sets. They make agents more consistent and useful by defining what they should do, when they should do it, and what they should avoid.
Learning note: These examples are for learning and guidance only. If a skill pattern is not useful in your workflow, ignore it and adapt the idea to your needs.
Example skill
---
name: aws-cost-optimizer
description: Use this skill when reviewing AWS cost, EKS usage, and Terraform resources for waste.
---
# AWS Cost Optimization Assistant
## Goal
Review infrastructure and suggest safe cost-saving changes.
## Instructions
1. Look for idle or oversized resources first.
2. Check EKS node groups, EC2 instances, EBS volumes, and NAT gateways.
3. Ignore ideas that reduce availability or break production.
4. Rank suggestions by impact and risk.
5. Return a short summary with evidence, risk level, and next action.
## Boundaries
- Do not suggest deleting production workloads without evidence.
- Do not recommend cost changes without checking current usage.
- Focus on actionable, low-risk recommendations.
Good skills include
- Clear trigger condition
- Specific objective
- Workflow steps
- Output format and guardrails
Common mistakes
- Too many responsibilities in one skill
- Missing constraints or limits
- No evaluation method
- Unclear trigger usage
Skill examples
Boundaries are as important as instructions
Pipeline troubleshooting skill
Goal: review CI failures quickly. Boundaries: check logs before giving a diagnosis, avoid guessing root cause, and prefer evidence-based steps.
Research skill
Goal: summarize a topic with references. Boundaries: do not invent sources, prefer cited evidence, and distinguish between fact and opinion.
Kubernetes incident skill
Goal: help find ingress or pod issues. Boundaries: do not suggest changing security policies without checking the current config, and ask for missing context when needed.