The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites via the capture_atlas, inspect_page, and audit_page tools. This presents a surface for indirect prompt injection where a malicious website could attempt to influence agent behavior through embedded instructions in its HTML or metadata.
Ingestion points: External URLs processed by capture and inspection tools in SKILL.md and v2-routing.md.
Boundary markers: The skill requires the agent to report coverage status and limitations, treating content as evidence rather than instructions.
Capability inventory: CLI execution (awwwards-index) and local file writes (export_evidence_pack).
Sanitization: Explicit redaction of credentials and query parameters from provenance; enforcement of robots.txt; and filtering of destination IP addresses.
[COMMAND_EXECUTION]: Instructions include using the awwwards-index CLI for cache maintenance tasks such as importing, reindexing, and pruning. These commands are local and scoped to the application's cache directory.
[EXTERNAL_DOWNLOADS]: The documentation references spectorjs@0.9.33 as an optional NPM dependency required for WebGL frame capture. This is a well-known utility used for debugging and its use is consistent with the skill's documented purpose.
[DATA_EXFILTRATION]: While the skill captures data from external URLs, it incorporates significant safety controls: it redacts credentials and sensitive query parameters from provenance, honors robots.txt, blocks private, loopback, and reserved IP ranges, and restricts data export to a fixed local directory within the cache.
[DYNAMIC_EXECUTION]: The skill involves the injection of a local spectorjs bundle into the browser context during scene capture to facilitate WebGL frame inspection. This is a documented feature of the vendor's toolset intended for design auditing.

awwwards-atlas