The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides explicit instructions for the agent to execute a bundled Node.js utility (scripts/eval_prompt_director.mjs). This script is used locally to initialize test environments, evaluate generated prompts, and produce markdown reports. It performs standard file system operations such as reading benchmark data and writing output to temporary directories (/tmp).
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for evaluating externally provided prompt files. This ingestion of local file content to drive the logic of the evaluator script constitutes an attack surface for untrusted data.
Ingestion points: Prompt files read from the local file system via the --outputs or --prompt-file arguments in scripts/eval_prompt_director.mjs.
Boundary markers: None; the content is processed as raw text.
Capability inventory: The script possesses the capability to read files (fs.readFileSync), write files (fs.writeFileSync), and create directories (fs.mkdirSync).
Sanitization: The script performs regex-based pattern matching on the ingested content to calculate scores rather than executing it or re-interpolating it into sensitive prompts.
[DYNAMIC_EXECUTION]: The evaluator script (scripts/eval_prompt_director.mjs) dynamically generates regular expressions at runtime using pattern strings defined in evals/benchmark-cases.json and internal configuration constants. This is used for automated scoring of prompt structural elements and capabilities.

gpt-image2-prompt-director