The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute standard filesystem and version control commands, including git diff, ls, mkdir, and grep. These are used appropriately for identifying modified assets and searching for text patterns within the local project environment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze content from CLAUDE.md, .claude/rules/, and other SKILL.md files which are composed primarily of natural language instructions. This creates a potential surface where adversarial content within those files could attempt to influence the agent's evaluation logic.
Ingestion points: Files identified in Step 1 (e.g., CLAUDE.md, .claude/rules/*.md, plugins/*/skills/*/SKILL.md) are read into the agent's context using the Read and Grep tools.
Boundary markers: None explicitly defined; the agent processes the raw content to perform static source inspection.
Capability inventory: The skill has access to Bash for command execution and Write for modifying the regression probe suite.
Sanitization: There is no explicit sanitization or escaping of the ingested data before the agent performs its inspection, though the logic is focused on pattern matching rather than direct execution of the ingested text.
[SAFE_PRACTICE]: The skill implements a mandatory human-in-the-loop (HITL) check in Step 6, requiring explicit user approval before using the Write tool to update baseline regression probes. This prevents the skill from autonomously altering its own test suite without oversight.

prompt-regression