The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains phrases often associated with prompt injection, such as "ignore previous instructions" and "repeat your system prompt". These are false positives. They are located in Section 1 (triggers), Section 7 (pitfalls), and Section 8 (safety notice) for the purpose of educating the agent on what to look for when auditing target code or how to handle adversarial inputs encountered during a review. The skill explicitly instructs the agent NOT to follow such commands.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze external codebases, which constitutes a potential attack surface for indirect prompt injection.
Ingestion points: Source code and configuration files provided via the [target-file-or-directory] argument.
Boundary markers: The skill instructions (LLM01) explicitly recommend the use of boundary markers and delimiters when building LLM pipelines.
Capability inventory: The skill is restricted to read-only tools (Read, Grep, Glob) as specified in the YAML frontmatter.
Sanitization: The skill provides logic for identifying missing sanitization and recommends specific remediation steps like using DOMPurify or parameterized queries.
[DATA_EXFILTRATION]: The skill does not contain any network-enabled tools, URLs to non-whitelisted domains, or instructions to transmit data externally. It is focused entirely on local static analysis.
[COMMAND_EXECUTION]: No shell commands, privileged operations (sudo), or script execution patterns were found. All analysis is performed using text-searching tools (Grep, Glob).

llm-top-10