The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains references to prompt injection patterns like 'ignore previous instructions' and 'system prompt extraction'. These are used as test objectives to verify that the target agent can properly handle such inputs, rather than being instructions to override the behavior of the agent running the skill. The skill also handles untrusted agent code (Indirect Prompt Injection surface). Ingestion points: Local agent code read via file tools. Boundary markers: Instructions to summarize findings into structured markdown (description.md) and YAML (risks.yaml). Capability inventory: Local file read/write and subprocess execution of the internal assembly script. Sanitization: The assembly script escapes scalar values using standard JSON encoding to prevent malicious data from breaking the YAML structure or injecting commands.
[COMMAND_EXECUTION]: The skill directs the agent to execute a provided Python script (scripts/build_scenarios.py) to validate and assemble scenarios. This script uses only standard library modules and performs deterministic data processing without dangerous operations or network access.
[DATA_EXFILTRATION]: The instructions emphasize local processing of the user's code and explicitly forbid uploading it to external services, ensuring that intellectual property is protected.
[EXTERNAL_DOWNLOADS]: The skill references the livekit-agents SDK and the lk CLI tool. These are official resources provided by the skill's author (LiveKit) and are necessary for the intended simulation functionality.

livekit-simulations