The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for LLM-as-a-judge evaluations, which creates a surface for indirect prompt injection where malicious content in evaluated data could influence the judge's scoring or reasoning.
Ingestion points: The skill ingests untrusted data via the input_text and output_text parameters in the Python SDK examples and the judge message content defined in SKILL.md.
Boundary markers: No specific delimiters or boundary instructions are provided to separate the evaluation criteria from the untrusted data being analyzed.
Capability inventory: The skill possesses the capability to modify LaunchDarkly project configurations, update targeting rules, and record metrics via REST API calls and SDK methods.
Sanitization: There are no instructions for sanitizing or filtering the content provided to the judges before evaluation.
[COMMAND_EXECUTION]: Provides shell commands using curl to interact with LaunchDarkly's beta REST API for creating configs, managing variations, and updating targeting instructions.
[EXTERNAL_DOWNLOADS]: References and provides links to official LaunchDarkly GitHub repositories and documentation for SDK implementation examples.

online-evals