The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses explicit override instructions: 'You are in AUTONOMOUS MODE. Do NOT ask questions. Do NOT pause for confirmation.' This attempts to bypass the agent's interactive safety protocols and decision-making logic.
[COMMAND_EXECUTION]: The instructions direct the agent to execute shell commands for validation purposes, including terraform validate, docker build --check, and kubectl dry-run. It also suggests the execution of docker compose to start the monitoring stack.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation of various instrumentation libraries from official registries, such as prom-client, prometheus_client, and OpenTelemetry packages. These are well-known observability tools.
[DATA_EXFILTRATION]: In the 'SELF-EVOLUTION TELEMETRY' section, the skill instructs the agent to write execution metadata to ~/.claude/projects/skill-telemetry.md. This targets a hidden directory in the user's home folder outside the standard project workspace.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface as it scans project files (YAML, JSON, Terraform, source code) and processes $ARGUMENTS to generate configuration and scripts.
Ingestion points: Reads project files during Infrastructure Detection (PHASE 1) and accepts user-provided arguments.
Boundary markers: None identified to separate untrusted file content from generation logic.
Capability inventory: Performs file writes, package installations, and suggests shell command execution.
Sanitization: Relies on a 'Self-Healing Validation' phase to check for secrets but lacks explicit sanitization for ingested file content.

monitoring