The Agent Skills Directory

[SAFE]: The skill establishes and maintains strict security boundaries through Rule 1 (Documentation only), which explicitly forbids the modification of executable code, build logic, or configuration values, and Rule 2 (Zero hallucination), which requires that all claims be grounded in source code read during the current session. The skill mandates the use of a [REDACTED] placeholder for any sensitive information such as passwords, tokens, or private keys discovered during the audit. Furthermore, the skill's reliance on specialized agents for claim verification and surface auditing ensures that all changes are factually accurate and do not introduce unintended behavior.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to its primary function of reading and processing untrusted data from an external codebase and pull request diffs. Ingestion points: The agent ingests data through the #codebase, #activePullRequest, and #changes references, resolving them via file-search and git tools as defined in SKILL.md. Boundary markers: Robust boundaries are established by the Zero Hallucination rule and the requirement for verbatim string proof, which forces the agent to treat external content as static data for verification rather than instructions to be executed. Capability inventory: The skill possesses capabilities for repository-wide file reading and documentation-file modification (creation, editing, and deletion of markdown files), but is strictly prohibited from executing code or modifying system behavior. Sanitization: The skill enforces mandatory credential redaction and requires every claim to be grounded in the implementation body, effectively preventing the agent from acting on malicious instructions embedded in the audited code.

audit-docs