The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted external content from the web via WebFetch and Browser automation, creating a surface for indirect prompt injection.
Ingestion points: WebFetch results, Browser page source, and API responses processed in SKILL.md.
Boundary markers: The skill does not explicitly instruct the agent to use delimiters when processing fetched content, though it does provide strict extraction prompts.
Capability inventory: Access to bash (curl, jq, python3), javascript_tool, and browser computer actions in SKILL.md.
Sanitization: Implements post-extraction cleaning (whitespace, HTML entities, unicode normalization) as defined in references/data-transforms.md.
[COMMAND_EXECUTION]: The skill uses standard command-line tools (curl, jq, python3) for data retrieval and processing in Phase 3. These operations are scoped to the primary purpose of web scraping and follow best practices like using /tmp/ for temporary storage.
[DYNAMIC_EXECUTION]: Utilizes javascript_tool within the browser context to perform precise data extraction and handle lazy-loaded content. This is a standard and expected functionality for browser-based scraping tools.
[METADATA_POISONING]: There is an inconsistency between the author stated in the developer context (sickn33) and the YAML frontmatter (renat). However, this appears to be a minor metadata discrepancy rather than a deceptive tactic intended to cause harm.

web-scraper