The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill documents the use of the !command`` syntax to execute shell commands at load time for gathering context. While this platform feature has a high privilege level, the skill's content consists of code block examples for educational purposes (e.g., git status, npm test) and does not invoke dangerous or exfiltrative commands within its own instructions.
[COMMAND_EXECUTION]: The skill provides numerous examples of automating workflows using the Bash tool, including gh pr view, kubectl cluster-info, and various build scripts. It emphasizes security best practices by teaching developers to use the allowed-tools frontmatter field to restrict command capabilities to the minimum required set.
[INDIRECT_PROMPT_INJECTION]: The skill describes a significant attack surface where commands ingest untrusted data from user arguments ($1, $ARGUMENTS) and external files (@path).
Ingestion points: User-supplied positional arguments and file inclusion markers.
Boundary markers: Not explicitly emphasized, relying on the prompt's logical structure.
Capability inventory: Commands are shown using Bash, Read, Write, and AskUserQuestion tools.
Sanitization: The skill provides high-quality mitigation examples, specifically demonstrating how to use shell commands like grep and test to validate and sanitize inputs before they are processed by the agent.
[SAFE]: Heuristic detections for destructive commands (such as dd and rm) in references/testing-strategies.md are false positives in this context. They are part of instructional material demonstrating how to create large files for testing purposes and how to clean up temporary test artifacts.

developing-commands