The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the discovery interview (user input) and external research (web search results) to generate learning plans, resources, and lessons. This creates a surface where malicious instructions in a research source or user input could influence the generated content.
Ingestion points: User input in discovery interview (SKILL.md); external high-trust sources found via online research (resources/course-generation.md).
Boundary markers: The skill requires citation of sources but lacks explicit delimiters or instructions to ignore embedded prompts within the researched content.
Capability inventory: File system write access for generating Markdown files and HTML lessons in the learning/ directory.
Sanitization: The skill includes robust instructions in resources/html-lessons.md to HTML-escape all content and validate href attributes against a whitelist (https and local relative paths) when generating HTML files.
[DYNAMIC_EXECUTION]: The skill generates interactive HTML lessons containing embedded Vanilla JS for recap quizzes.
Evidence: The resources/html-lessons.md file defines a template for generating .html files that include <script> blocks or event-driven logic (e.g., reveal-on-click answers).
Mitigation: The skill specifically instructs the agent to use textContent instead of innerHTML and to escape all inserted text to prevent arbitrary code injection into the generated lessons.

teach-me