The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and process untrusted content from the public web, which may contain adversarial instructions.
Ingestion points: SKILL.md Step 2 ("Research what is missing") specifically directs the agent to gather data from reviews, forums, Reddit, and X.
Boundary markers: Absent. While the skill mandates verbatim quotes and source linking, it lacks instructions for the agent to ignore or neutralize embedded directives within the fetched content.
Capability inventory: The skill possesses file-writing capabilities (docs/product/proposal.md), network access for fetching external URLs, and the ability to generate interactive HTML code.
Sanitization: The skill does not define sanitization procedures to prevent the agent from obeying instructions hidden in research data.
[DYNAMIC_EXECUTION]: The skill generates interactive HTML code for prototypes based on external research data.
Evidence: SKILL.md Step 6 ("Build what the owner reviews, interface first") instructs the agent to create interactive HTML prototypes using "realistic sample data."
Risk: If the "realistic sample data" includes content fetched from external sources in Step 2 without proper escaping or sanitization, it could lead to the execution of malicious scripts (e.g., Cross-Site Scripting) when the user opens the prototype in a browser.

propose