The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from local git diffs which may contain hidden instructions for the agent. Evidence: Ingestion occurs during diff scanning in SKILL.md. Boundary markers are used in the reporting format in references/review-workflow.md, but lack explicit instructions to disregard data content. Capabilities include modification of source files and execution of verification suites. No sanitization of the scanned code is specified.
[COMMAND_EXECUTION]: The skill directs the agent to run shell commands like git diff to identify changes and suggests executing project-specific verification tools (tests, linters) after applying edits, as described in SKILL.md.

deslop