The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface where untrusted data from a Kubernetes cluster (such as log lines or resource metadata) could be pasted into the conversation. It mitigates this by requiring the user to use a wrap helper function.
Ingestion points: User-pasted output from kubectl get pods, kubectl logs, and helm get values (referenced in SKILL.md and references/untrusted-output.md).
Boundary markers: The skill uses a nonce-delimited <untrusted-data> tag system to isolate external data from instructions.
Capability inventory: The skill guides the user through helm and kubectl operations including upgrade --install, apply, create, patch, and annotate.
Sanitization: The wrap function includes sed and tr commands to strip ANSI escape sequences and non-printable control characters that could be used to manipulate terminal display.
[EXTERNAL_DOWNLOADS]: The skill references and downloads resources from official vendor-controlled and well-known domains.
Evidence: Downloads Helm charts from https://observeinc.github.io/helm-charts and fetches configuration guidance from https://docs.observeinc.com.
Evidence: References the official open-telemetry/opentelemetry-operator for EKS Fargate deployments.
[COMMAND_EXECUTION]: The skill provides various Kubernetes management commands for the user to execute.
Evidence: Generates helm upgrade --install and kubectl apply commands for cluster configuration.
Safety Measure: Explicitly instructs the agent not to run these commands itself and provides a method to use sensitive tokens ($OBSERVE_TOKEN) without them ever entering the AI's context.

setup-k8s-collection