Agent Skills: Operate Infrastructure

Operate Infrastructure

Boundary

• Work from files, plans, logs, and read-only commands before changing anything.
• Do not run apply, delete, destroy, or rollback until identity, exact resources, blast radius, and plan/diff/inventory are shown and the user confirms.
• If the task is deployment, rollout, rollback, or production apply, use deploying-infra.
• If the task is only shell scripts, generic command pipelines, or the shell body inside a GitHub Actions run: step, use writing-shell.
• For GitHub Actions, workflow structure, triggers, jobs, permissions, runners, actions, environments, secrets, caching, concurrency, and policy stay here. Mixed workflow and shell-body changes compose with writing-shell.

Role behavior

• Write-capable: make minimal file changes and run safe validation. Stop before live mutation unless the user confirmed exact resources.
• Read-only: apply nothing; return proposed file changes, evidence, and validation commands.

Load references

Load every matching reference:

• Terraform/OpenTofu files, modules, state, or plans → terraform.md
• Kubernetes manifests or kustomization.yaml → kubernetes.md
• Chart.yaml, Helm values, or chart templates → helm.md
• GitHub workflow YAML outside pure run: shell bodies → github-actions.md
• Dockerfile or container image build/release concerns → dockerfile.md
• AWS CLI, EC2, ECS, Lambda, S3, RDS, IAM, or CloudWatch → aws.md
• GCP CLI, GCS, Compute Engine, IAM, quotas, or Cloud Logging → gcp.md
• Cloud Run services, revisions, traffic, or logs → cloud-run.md
• BigQuery queries, tables, datasets, or cost checks → bigquery.md
• Linux services, hosts, processes, disks, or networks → linux.md

Mixed stacks: load all matching references. Unknown stack: use the workflow below only.

Workflow

1. Identify scope: files, resources, environment, account/project, region/zone, and owner.
2. Verify cloud identity before cloud work; prefer explicit profile/project/region over defaults.
3. Inspect current state with read-only evidence: files, plan/diff, list/describe/status, logs, metrics, and recent events.
4. For authoring/design: choose the smallest pattern that preserves ownership, state boundaries, and least privilege.
5. For troubleshooting: rank likely causes, gather one safe signal, then propose the next step.
6. For validation: run relevant gates when tools exist; state skipped gates and why.
7. For destructive, costly, or externally visible work: show exact resources and blast radius, then stop for confirmation or hand off to deploying-infra.

Validation gates

• Terraform/OpenTofu: format, init without backend when possible, validate, plan, tflint, checkov or trivy config; use plan JSON for policy checks when needed.
• Kubernetes/Kustomize: render first, schema-check with kubeconform, then policy/security-check with kube-linter, kubescape, conftest, or kyverno.
• Helm: lint chart, render templates, use helm diff before upgrade planning, validate rendered YAML.
• Docker/images: lint Dockerfile with hadolint; scan images/config with trivy; use syft, grype, and cosign where SBOM, vulnerability, or signature proof matters.
• GitHub Actions: run actionlint and zizmor; require SHA-pinned actions and least-permission jobs.
• Cloud CLI: verify identity, inventory resources, estimate cost or dry-run when available, and check IAM/quota before mutation.

Output

INFRA RESULT
============
Scope: <files/resources/environment>
Identity: <account/project/profile/region or not applicable>
Status: DONE | NEEDS CONFIRMATION | BLOCKED

Evidence:
- <file:line, plan/log/status summary, or command result>

Changes or proposal:
- <minimal change or proposed next step>

Validation:
- <gate> — pass/fail/skipped

Next:
- <safe next action, confirmation request, or none>