The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a data-driven optimization loop that ingests untrusted LLM outputs and production traces to generate revised system prompts. This creates a surface for indirect prompt injection where malicious instructions embedded in production data could influence the agent's prompt generation logic.
Ingestion points: Production trace data is retrieved via ax spans export, and experiment results are gathered via ax datasets export and ax experiments export.
Boundary markers: The meta-prompt template in references/optimization-meta-prompt.md uses structured headers such as ORIGINAL BASELINE PROMPT and PERFORMANCE DATA to separate instructions from data, though these are not cryptographically secure.
Capability inventory: The skill utilizes the ax CLI for data retrieval, jq for JSON processing, and shell redirection to write trace data to local files (e.g., spans.json).
Sanitization: The instructions do not specify sanitization or escaping of the ingested trace data before it is interpolated into the optimization template.
[COMMAND_EXECUTION]: The skill frequently executes shell commands, specifically the ax CLI tool and jq, to manage Arize profiles and process production data. These are standard operations for the skill's purpose.
[DATA_EXFILTRATION]: The skill exports production spans and experiment results to the local filesystem. This is a primary function for prompt debugging and optimization within the Arize ecosystem.

arize-prompt-optimization