The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill references a specific absolute file path on the host system (/Users/engineer/workspace/backtest/.claude/workflows/pairwise-eval-workflow.js) and instructs the agent to utilize this JavaScript file for handling parallel task execution and subagent fan-outs.- [DYNAMIC_EXECUTION]: Step 3 includes a Python script designed to perform a 'leak-check' by reading files from the local filesystem and matching their content against variant names using regular expressions. The skill expects the agent to generate and execute this script locally to ensure blinding integrity.- [INDIRECT_PROMPT_INJECTION]: The workflow relies on a judging agent to process and evaluate untrusted data generated by other AI variants. This creates a surface for indirect prompt injection, where a variant could produce adversarial text intended to influence the judge's final verdict.
Ingestion points: The judge subagent reads and processes files from the blind/ directory (e.g., INPUT-A.md and INPUT-B.md) as described in Step 4.
Boundary markers: The judge prompt includes instructions to follow a rubric and apply anti-gaming rules, which serves as a functional boundary, though no explicit technical delimiters are applied to the input content.
Capability inventory: The judging subagent is tasked with reading files and producing a text report; it is not explicitly granted tool access or sensitive capabilities in the provided judge prompt.
Sanitization: The skill provides a mechanism to check for leaks of variant names, but does not include any filtering or sanitization to detect or neutralize prompt injection sequences within the variants' outputs.

ai-evaluate