The Agent Skills Directory

[SAFE]: The skill facilitates skill management and evaluation using structured workflows and a dedicated Python executor script.
[COMMAND_EXECUTION]: The scripts/eval-executor.py script executes the claude CLI and git commands to run evaluations and manage worktrees. Execution is performed using list-based arguments without a shell, which is a standard security practice to prevent command injection. The tool also provides functionality to fingerprint directories to ensure that only approved skill versions are executed.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection as it ingests user-provided 'held-out test requests' and executes them verbatim via the claude CLI.
Ingestion points: User input collected during Phase A dialog in references/create.md (Round 3) and references/update.md (Round 2).
Boundary markers: Absent; requests are passed directly to the executor to test raw agent response behavior.
Capability inventory: scripts/eval-executor.py invokes the claude CLI for task execution, capturing output and metadata across multiple trials.
Sanitization: Absent; the skill relies on the execution environment's tool restrictions (--allowedTools) and the underlying agent's own safety guardrails.

recipe-eval-skill