The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents how to define steps that execute shell commands (type: shell) and binary executions (type: exec) within the Atmos framework. This is the intended core functionality of the tool for infrastructure orchestration.
[DYNAMIC_EXECUTION]: The skill provides guidelines for the type: script step, which supports various interpreters such as Python, Bash, and Starlark. It specifically notes that remote sources are not supported for script files, which is a security-positive constraint.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential injection surface where Atmos renders template actions {{ ... }} inside script bodies. It explicitly instructs users and the agent to mitigate this risk by using context objects (ctx.flags) or using the !literal tag to prevent unintended interpolation of untrusted data.
[DATA_EXFILTRATION]: The skill documents the http (and webhook alias) step type, which allows for outbound network requests. While this could theoretically be used for exfiltration, it is documented as a standard feature for CI/CD notifications and integrations.
[SAFE]: The skill is authored by the official vendor (Cloud Posse) and adheres to established security best practices for the tool being documented. No obfuscation, prompt injection, or hidden persistence mechanisms were detected.

atmos-steps