Working with AWS

Persistent guardrails that apply to every AWS task. Load this first; the other skills in this plugin assume these rules are already in effect.

Core rules

• Prefer the AWS MCP Server. Use it for AWS interactions — it provides sandboxed execution, CloudWatch metrics, CloudTrail audit logging, and IAM-based access control. If it is unavailable, fall back to the AWS CLI directly. See the aws-mcp-server skill for how to drive it.
• Discover skills before starting. Before beginning a task, check whether a relevant AWS skill exists (via the MCP server's retrieve_skill, or the skills bundled here). Load it and prefer its guidance over general knowledge.
• Verify, don't guess. When uncertain about API parameters, permissions, quotas, or error codes, confirm against current AWS documentation (search_documentation / read_documentation) rather than relying on training data that may be stale. State uncertainty explicitly if you cannot confirm.
• Infrastructure-as-code first. When creating infrastructure, produce AWS CDK or CloudFormation/SAM rather than issuing direct mutating CLI commands. IaC is reviewable, repeatable, and safely destroyable. Reserve direct CLI/API calls for reads, diagnostics, and one-off inspection.
• Follow Well-Architected. Apply the six pillars — Operational Excellence, Security, Reliability, Performance Efficiency, Cost Optimization, and Sustainability. Call out the tradeoff when a request conflicts with a pillar.
• No em dashes in resource names or descriptions. Use hyphens instead.

Security defaults

• Least privilege. Never provision or assume admin permissions when a scoped role will do. Grant only the actions and resources a workload needs.
• Read-only in production. Prefer IAM policies that let agents suggest but not execute mutating actions in production. The AWS MCP Server exposes the aws:ViaAWSMCPService and aws:CalledViaAWSMCP condition keys so you can write policies (and SCPs) that apply specifically to agent-initiated calls.
• Private by default. Block public access on new storage, encrypt data at rest and in transit, and enforce TLS.

Secret safety

• For any secret, credential, API key, token, or password task, do not print secret values into context or logs.
• Do not call secretsmanager get-secret-value / batch-get-secret-value just to read a value into the conversation. Reference secrets by name/ARN, and resolve them at runtime — e.g. CloudFormation dynamic references like {{resolve:secretsmanager:secret-id:SecretString:json-key}} — so the value never enters the agent context.
• Never commit credentials to source, IaC parameters, or plugin configuration.

Cost awareness

• Prefer scale-to-zero and on-demand options for prototypes and demos.
• Flag any resource with a standing hourly floor cost (NAT gateways, RDS/Aurora provisioned instances, always-on load balancers) before creating it.
• Tag resources for cost allocation where the account uses tagging conventions.