The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates development and validation workflows by executing shell commands through the Bash tool and custom Python scripts. These commands facilitate repository management, index generation, and the invocation of external CLI tools like the claude and Go toolchains.
[DYNAMIC_EXECUTION]: Several scripts in the scripts/ directory, including run_eval.py, optimize_description.py, and eval_compare.py, use the subprocess module to execute external binaries at runtime. While these scripts use list-based arguments to mitigate shell injection, they directly incorporate user-supplied prompts and test queries into the command-line arguments passed to the claude CLI.
[INDIRECT_PROMPT_INJECTION]: The skill's evaluation architecture ingests untrusted data in the form of test prompts provided by the user. These prompts are subsequently processed by the skill and evaluated by sub-agents (grader.md, comparator.md), creating a vector where malicious instructions hidden in test data could influence the outcome of the evaluation or benchmarking process.
Ingestion points: User-provided test prompts are ingested via scripts/run_eval.py and scripts/optimize_description.py (via trigger-eval.json).
Boundary markers: The skill does not implement explicit delimiters or safety markers to isolate untrusted test content during the evaluation runs.
Capability inventory: The skill is highly capable, with the ability to read and write files, execute shell commands, and orchestrate multiple agent sessions.
Sanitization: The workflow relies on manual human review and quantitative assertions rather than automated prompt sanitization to detect or prevent injection in test data.
[SAFE]: The skill includes strong security best practices in its documentation, such as explicit warnings against hardcoding secrets (found in references/error-catalog.md) and a mandatory path placeholder convention (ADR-201) to prevent the leakage of developer-specific environment paths in shared documentation.

skill-creator