The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reviews untrusted source code and diffs which may contain adversarial instructions.
Ingestion points: The skill ingests staged and unstaged changes, branch diffs, and surrounding source code (SKILL.md).
Boundary markers: The skill explicitly warns against prompt injection and directs the agent to treat untrusted content as user-role data rather than instructions (SKILL.md).
Capability inventory: The agent can execute shell commands like grep, npm audit, sqlfluff, and tsc, modify local files (with user approval), and run PREPARE queries on a local database (SKILL.md).
Sanitization: The instructions guide the agent to verify sanitization in the reviewed code and use role-based separation for its own input processing.
[COMMAND_EXECUTION]: The skill utilizes several developer tools and linters to verify code quality and security.
Evidence: Specifically mentions running grep, npm audit, cargo audit, pip-audit, tsc, knip, sqlfluff, pgsql-parser, and vulture (SKILL.md).
[DYNAMIC_EXECUTION]: The skill recommends validating embedded strings against real engines to ensure syntax correctness.
Evidence: Recommends executing PREPARE statements on a local database, compiling regex in a REPL, and parsing GraphQL using project libraries (SKILL.md).

dev-review