The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a system that generates code based on external inputs (B-UML models).
Ingestion points: Generators ingest DomainModel, GUIModel, and Agent models from files like SKILL.md and references/api-and-web.md.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are documented for the generated code.
Capability inventory: The skill documents file system writes (creation of output/ directories and source files) and subprocess execution (DjangoGenerator and SQLGenerator).
Sanitization: The SKILL.md instructions explicitly note that method bodies for BAL or CODE implementation types are inserted verbatim into generated handlers, which could allow malicious code injection if the source model is attacker-controlled.
[COMMAND_EXECUTION]: Several generators utilize shell commands or subprocesses to complete their tasks.
DjangoGenerator in references/api-and-web.md runs django-admin startproject and manage.py startapp.
SQLGenerator in references/persistence.md executes a Python subprocess to generate DDL artifacts.
[DYNAMIC_EXECUTION]: The skill facilitates the creation of executable content from model definitions.
The system allows users to provide raw code strings (MethodImplementationType.CODE) that are compiled or interpreted in the generated application.
PydanticGenerator dynamically translates OCL constraints into Python validator logic.

besser-generators