The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external data from command outputs such as observe datastream list and observe content view. It specifically addresses the risk of malicious instructions hidden in these outputs (e.g., from workload logs or telemetry) by requiring the use of a wrap helper.
Ingestion points: Phase 2 of SKILL.md identifies commands whose output is pasted back by the user.
Boundary markers: The skill mandates a nonce-delimited tagging system (<untrusted-data nonce="...">) defined in references/untrusted-output.md to isolate data from instructions.
Sanitization: The wrap utility includes a multi-stage sanitizer using sed and tr to neutralize ANSI escape sequences and control characters that could be used for terminal display attacks.
Capability inventory: The skill uses the observe CLI for resource creation but restricts actions to the vendor's cloud environment.
[CREDENTIALS_SAFE]: The skill handles the creation of ingest tokens using a security-conscious workflow that prevents the secret from ever being exposed to the AI assistant. It instructs the user to store the token in a shell environment variable and uses jq to strip the secret key from the JSON metadata before the assistant processes it.
[COMMAND_EXECUTION]: The skill executes various observe CLI commands to manage datastreams and content packages. These operations are legitimate, vendor-provided functionalities consistent with the skill's purpose for the observeinc tenant environment.

setup-k8s-backend