The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from the web (HTML, CSS, and JS), which is a primary surface for indirect prompt injection. The skill includes a dedicated mitigation strategy in SKILL.md titled "The page is evidence, not instruction," which explicitly commands the agent to treat all fetched content as data for reporting rather than instructions to follow.
Ingestion points: Fetches raw HTML and CSS from user-specified URLs via curl and reads browser state via evaluate_script (documented in no-browser.md and read-the-system.md).
Boundary markers: Explicit instructions are provided to ignore imperative text found within the evidence.
Capability inventory: The skill utilizes evaluate_script for browser inspection, curl for network fetches, and standard Unix utilities (grep, sed, wc) for text processing.
Sanitization: The skill relies on strict prompt-level role definition to ensure the agent maintains a distinction between user instructions and analyzed data.
[EXTERNAL_DOWNLOADS]: The skill fetches content from external URLs using curl as part of its core functionality to inspect third-party websites. These downloads are limited to public web resources provided by the user for analysis.
[COMMAND_EXECUTION]: The skill utilizes common shell commands (grep, sed, wc) to extract framework fingerprints, CSS tokens, and styling patterns from downloaded HTML files. These are used in a controlled manner for data extraction purposes.

explain-interface