The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided brand content (e.g., website copy, blog posts) to influence code generation, creating a vulnerability surface where malicious instructions in the input could affect the output.
Ingestion points: Step 2 of the Generation Workflow in SKILL.md instructs the agent to collect and analyze existing brand content from the user.
Boundary markers: No boundary markers, delimiters, or "ignore embedded instructions" warnings are defined to isolate untrusted user data during processing.
Capability inventory: The skill utilizes scripts/landing_page_scaffolder.py to generate and potentially write TSX/HTML files to the file system and invokes brand_voice_analyzer.py for content analysis.
Sanitization: The scripts/landing_page_scaffolder.py script lacks sanitization for user-controlled strings when generating Next.js/React TSX components, allowing untrusted content to be directly embedded in the generated code.
[DYNAMIC_EXECUTION]: The skill uses a Python script (scripts/landing_page_scaffolder.py) to generate executable React/Next.js (TSX) code. The script utilizes string interpolation (f-strings) to assemble components from configuration data. While the script performs HTML escaping for the standalone HTML format, it does not sanitize user-controlled fields (like headlines, taglines, or descriptions) for the TSX format, which could result in the generation of components containing malicious scripts if the agent is manipulated by indirect injection.
[COMMAND_EXECUTION]: The instructions in SKILL.md direct the agent to execute an external script (brand_voice_analyzer.py) located at a path outside the skill's own package (marketing-skill/skills/content-production/scripts/). This introduces a dependency on the host environment's file system structure and the security of scripts not included in this skill's distribution.

landing-page-generator