The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to audit external websites by fetching their content, which presents a significant untrusted data ingestion surface.
Ingestion points: Website HTML and metadata are fetched and processed via src/crawler/fetcher.ts and src/crawler/playwright-fetcher.ts.
Boundary markers: The skill applies a sophisticated "Trust Model" documented in SKILL.md and implemented in src/reporters/llm-reporter.ts. This system wraps all site-derived content in unique, per-report nonce-stamped delimiters (<untrusted-{nonce}>) and prepends a <security-notice> instructing the agent to treat that content as data only.
Capability inventory: The skill utilizes the Bash(seomator:*) tool to run the local audit engine.
Sanitization: The reporter automatically strips invisible characters used for prompt injection, including zero-width characters (U+200B–U+200D, etc.) and the Unicode tag block (U+E0000–U+E007F), before outputting data to the agent.
[OBFUSCATION]: A static analysis detector flagged the presence of hidden Unicode characters in src/reporters/llm-reporter.test.ts.
Evidence: The file contains strings such as visible\u{E0049}\u{E006E}\u{E0073}\u{E0074}\u{E0072}.
Verification: These characters are contained within a unit test suite. They are used exclusively to verify that the tool's security filters successfully detect and strip invisible prompt-injection vectors. The code is benign and represents a high standard of security testing.

seo-audit