The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands and scripts (ralph, uv run python, chmod, mkdir) to scaffold and validate project directories. These operations are scoped to the intended function of the skill and use restricted tool prefixes.
[SAFE]: The included validation script (scripts/validate.py) enforces strict security policies, such as using yaml.safe_load and explicitly rejecting shell metacharacters (e.g., |, &&, ;, backticks) in command strings to prevent shell injection.
[SAFE]: The skill architecture provides a defense-in-depth approach to Indirect Prompt Injection. It documents that placeholders are resolved in a single pass, preventing untrusted data from external sources (like command outputs) from being re-expanded or interpreted as prompt instructions.
[SAFE]: Multiple 'Guard' mechanisms are implemented to ensure autonomous safety, including mandatory iteration caps enforced by a wrapper script (scripts/run.sh) and explicit success sentinels to prevent runaway loops or silent failures.

ralphify-spec