The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The tool's installation instructions recommend piping a remote shell script from the author's GitHub repository directly into a shell interpreter (curl | sh and irm | iex). This pattern executes remote code without local verification of the script content.
[COMMAND_EXECUTION]: The evaluation runner (scripts/run_evals_template.py) and various scripts within the skill's lifecycle execute arbitrary shell commands defined in JSON evaluation specifications using subprocess.run with shell=True.
[DYNAMIC_EXECUTION]: The skill uses dynamic execution patterns for validation and file generation. Specifically, scripts/check_pipeline.py uses the Python compile() function to check the syntax of generated scripts, and scripts/render_installers.py dynamically creates installer scripts by performing string replacements on templates and modifying file permissions.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a factory that ingests untrusted data from documents, emails, and transcripts to automatically derive intent and generate new skill code and instructions. While it includes a security scanner for the skills it produces, the ingestion process itself is an attack surface.
[EXTERNAL_DOWNLOADS]: Scripts such as scripts/dependency_health.py and scripts/schema_drift.py use urllib.request.urlopen to perform network requests to external URLs for monitoring. The judge harness also makes outbound calls to the Anthropic API.

agent-skill-creator