The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides scripts and instructions that execute shell commands such as npm install, npm run build, zip, and curl to manage the build and deployment workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user project files and build configurations, creating a potential surface for indirect prompt injection.
Ingestion points: User project directories and build command strings provided in the examples within SKILL.md and README.md.
Boundary markers: None provided to separate agent instructions from untrusted data.
Capability inventory: Shell command execution via subprocess.run (Python) and shell scripts, network requests via curl and requests, and file system operations.
Sanitization: The skill does not implement validation or sanitization for the build commands or project content before execution.
[DYNAMIC_EXECUTION]: The Python deployment example in SKILL.md uses subprocess.run(build_cmd.split(), ...) to execute commands. This pattern allows for the execution of arbitrary commands if the build_cmd parameter is influenced by untrusted input.
[EXTERNAL_DOWNLOADS]: The skill uses npm to fetch dependencies and curl/requests to communicate with api.rebyte.ai and storage.googleapis.com. These operations target vendor-specific or well-known cloud services for legitimate deployment functionality.

web-app-builder