The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides numerous templates (e.g., data_processing.sh and content_pipeline.sh) that ingest untrusted external data and interpolate it directly into LLM prompts without sanitization or boundary markers.
Ingestion points: Data enters the context via cat $file in the processing script and from API responses stored in the RESEARCH variable.
Boundary markers: The templates do not use delimiters or instructions to ignore embedded commands within the processed data.
Capability inventory: The skill utilizes Bash(belt *) for shell access, subprocess.run in Python scripts, and has network capabilities via curl and the belt CLI.
Sanitization: No escaping, filtering, or validation is performed on the data before it is interpolated into the JSON-formatted prompts.
[PERSISTENCE]: The skill explicitly documents how to establish persistence on the host system using cron jobs to automate AI workflows.
Evidence: The 'Scheduled Automation' section provides instructions for using crontab -e to set up daily, weekly, and hourly task execution.
[EXTERNAL_DOWNLOADS]: The skill encourages the installation of additional skills and CLI tools from remote sources.
Evidence: Instructions include npx skills add belt-sh/cli and multiple references to other skills hosted in the inference-sh GitHub repository.
[COMMAND_EXECUTION]: The skill relies on extensive shell scripting and Python's subprocess module to orchestrate external tools.
Evidence: The automation.py script uses subprocess.run to execute the belt CLI with dynamic input data, and multiple bash scripts perform file system operations and network calls.

ai-automation-workflows