The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions frequently reference the use of mcp__claude-in-chrome__javascript_tool with the javascript_exec action. This involves executing JavaScript code that is dynamically generated or assembled at runtime to query the DOM of external web pages. While this is the intended functionality of the browser tool, it represents a dynamic execution pattern where untrusted page structures influence the generated code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and process data from arbitrary external web pages, creating an indirect prompt injection attack surface.
Ingestion points: Untrusted data enters the context via the read_page, get_page_text, and javascript_tool tools as described in Phase 4 and Phase 5 of SKILL.md.
Boundary markers: The skill mandates a planning phase using sequential-thinking to define a schema, which helps constrain the agent's focus, but it does not provide explicit instructions to wrap or delimit the ingested web content with safety warnings.
Capability inventory: The skill utilizes browser execution via javascript_tool and data persistence via memory_store (Phase 7).
Sanitization: The skill includes logic for escaping characters when transforming data to CSV or JSON formats, but it does not specify sanitization or filtering of the raw content to prevent embedded instructions from influencing the agent.
[METADATA_POISONING]: The author identified in the skill metadata (ruv) does not match the provided author context (dnyoussef). This discrepancy in the author field in SKILL.md is a metadata inconsistency, although it is likely an artifact of using a shared template.

web-scraping