The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from AI application traces (queries, responses, and retrieved context) for evaluation purposes, creating a vulnerability surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context via functions like evaluate_dataframe in references/evaluate-dataframe-python.md and runExperiment in references/experiments-running-typescript.md.
Boundary markers: The skill documentation (e.g., in references/evaluators-custom-templates.md) recommends using XML tags like <context>{{context}}</context> as delimiters to reduce the risk of the model following instructions embedded in the data.
Capability inventory: The skill possesses the capability to invoke LLMs for judging and execute custom Python or TypeScript code through its code-based evaluator framework.
Sanitization: The skill provides a framework for interpolation but does not provide built-in automatic escaping or sanitization of the untrusted data; this is left to the user's implementation of prompt templates.
[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions to download and install several external dependencies.
Python packages: Fetches arize-phoenix, arize-phoenix-client, arize-phoenix-evals, and arize-phoenix-otel which are the vendor's own official libraries.
TypeScript packages: Fetches @arizeai/phoenix-client, @arizeai/phoenix-evals, and @arizeai/phoenix-otel from the vendor's official NPM scope.
Well-known services: Recommends installing standard SDKs from trusted providers including openai, anthropic, and google-genai to support judge models.

phoenix-evals