The Agent Skills Directory

[PROMPT_INJECTION]: The skill suite possesses a substantial surface for indirect prompt injection attacks. Numerous specialist skills are designed to ingest and process untrusted external data—such as website HTML, marketing drafts, and analytics data—which the AI then interprets to generate reports, scores, or new content.
Ingestion points: Various scripts like schema_validator.py (HTML), content_scorer.py (Markdown), humanizer_scorer.py (Markdown), and attribution_analyzer.py (JSON) are designed to read external files.
Capability inventory: The AI agent is authorized to execute a variety of local Python scripts to manipulate this data and create files based on the results.
Risk: Malicious instructions hidden within the data files (e.g., in HTML comments or markdown text) could attempt to influence the agent's behavior during the analysis phase.
[COMMAND_EXECUTION]: The documentation for the prompt-engineer-toolkit describes a dynamic execution pattern involving the prompt_tester.py script. This tool accepts a --runner-cmd argument that serves as a template for shell commands, substituting {prompt} and {input} placeholders at runtime.
Evidence: Found in prompt-engineer-toolkit/SKILL.md which instructs the use of: python3 scripts/prompt_tester.py --runner-cmd 'my-llm-cli --prompt {prompt} --input {input}'.
Risk: This pattern of dynamically assembling shell commands from templates can lead to arbitrary command injection if the variables (sourced from potentially untrusted prompt or test-case files) are not rigorously sanitized before the shell executes the resulting string.

marketing-skills