The Agent Skills Directory

[SAFE]: No malicious patterns or security vulnerabilities were found.
[INDIRECT_PROMPT_INJECTION]: The skill correctly identifies that processing user-provided 'briefs' and existing repository files creates an indirect prompt injection attack surface. It addresses this by providing mandatory 'Agentic setup' instructions that define an untrusted content boundary, requiring the agent to treat external input as data rather than instructions. The skill also enforces regex-based validation for all user-supplied strings before they are used in shell commands or file operations.
[PROMPT_INJECTION]: Detection patterns for 'ignore previous instructions' were found in references/agentic-setup.md and references/shared-boilerplate.md. These findings are confirmed as false positives; the text consists of security rules that instruct the agent to detect and ignore such injections in untrusted data.
[COMMAND_EXECUTION]: The skill uses local command-line tools (git, bash, awk, grep, sed, diff) to perform repository maintenance tasks like linting and verifying file integrity. These operations are limited to the local environment and are protected by input sanitization logic.

om-create-skill