The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project files including source code, manifests, and git history, creating a surface for indirect prompt injection. Malicious instructions embedded in comments or documentation within the audited project could attempt to influence the reviewer agents or suppress findings. The skill mitigates this by instructing agents to treat such content as untrusted and requiring human-verifiable reasons for dismissing findings. Ingestion points: Project source files, configuration manifests (e.g., package.json), and git history. Boundary markers: Reviewer prompts explicitly define repository content as untrusted input. Capability inventory: Modifying local VCS-tracked source files, writing state to .outline/audit/, and executing local verification tools. Sanitization: Findings are normalized and consolidated through an adjudication contract requiring descriptive reasons.
[DYNAMIC_EXECUTION]: The skill generates and applies "fix batches" to source code based on AI suggestions. While these are applied to VCS-tracked files and verified by local test suites, the automated modification of executable code constitutes a dynamic execution risk.
[COMMAND_EXECUTION]: The skill executes local shell commands including git (ls-files, log, restore, revert), ast-grep, and various project-specific test runners (npm test, pytest, cargo test). These are used to resolve scope, analyze code churn, and verify suggested fixes within the skill's defined authority.

audit-project