The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides script templates that interpolate untrusted external data directly into AI prompts, creating a surface for potential injection attacks.
Ingestion points: Script examples read untrusted data from command-line arguments (e.g., $1 in conditional_workflow.sh) and from local files (e.g., $(cat $file) in data_processing.sh).
Boundary markers: No delimiters or "ignore embedded instructions" warnings are used when inserting external content into the model prompts.
Capability inventory: The workflows can execute AI model runs via the belt CLI and perform filesystem writes via shell redirection.
Sanitization: No input validation or escaping is performed on the ingested data before prompt interpolation.
[PERSISTENCE]: The skill explicitly documents how to set up scheduled automation using crontab entries. While this is a legitimate feature for the stated purpose of AI automation, it is a mechanism used to maintain execution across system restarts.
[COMMAND_EXECUTION]: The provided templates include numerous shell scripts and a Python script using subprocess.run to execute commands. These scripts perform operations such as creating directories (mkdir -p), reading files, and running the belt CLI tool.
[EXTERNAL_DOWNLOADS]: The skill references and suggests downloading code and configuration from external sources.
References installation instructions hosted on a GitHub repository for the inference-sh platform.
Suggests installing additional skills via npx skills add from repositories associated with the platform's ecosystem.
Includes an example script using curl to send data to a placeholder webhook domain (your-webhook.com) for error monitoring.

ai-automation-workflows