The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content which could contain malicious instructions designed to influence agent behavior.
Ingestion points: download.py fetches HTML from arbitrary target URLs, and references/explore-site.md facilitates discovery of new URLs.
Boundary markers: The skill contains strict hard constraints (in SKILL.md) prohibiting the main agent from reading, grepping, or parsing HTML files directly. It mandates the use of a subagent (/scrape-analyze-page) for extraction.
Capability inventory: The skill executes local Python scripts (uv run), writes validation specs to the filesystem, starts a local HTTP server (run_review.py), and invokes subagents.
Sanitization: The architecture uses subagent isolation to ensure that raw HTML never enters the main agent's context. The main agent only consumes structured JSON extraction results.
[COMMAND_EXECUTION]: The skill performs a runtime command to prepare the environment.
Evidence: download.py executes subprocess.run([sys.executable, "-m", "playwright", "install", "chromium"]) to ensure the Chromium browser is available for rendering.
Note: This is a standard setup operation for the Playwright library, which is a well-known service maintained by Microsoft.
[EXTERNAL_DOWNLOADS]: The skill is primarily intended to download external resources.
Evidence: download.py uses Scrapy and Playwright to fetch web pages. It also triggers the download of the Chromium browser binary if missing.
Note: These downloads are core to the skill's stated purpose and utilize established libraries and official browser registries.

scrape-plan