The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from local configuration files and interpolates it into prompts for an image generation API.
Ingestion points: Data is read from the style, style_description, and description fields within projects/{project_name}/project.json (as implemented in scripts/generate_clue.py).
Boundary markers: The prompt templates in SKILL.md lack delimiters or instructions to prevent the model from following instructions embedded within the injected project data.
Capability inventory: The skill is capable of performing local file writes to save image assets and updating project metadata files through the ProjectManager class.
Sanitization: The skill does not perform validation or sanitization of the strings retrieved from project files before they are used in the workflow.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute a local Python script (scripts/generate_clue.py) using shell commands.
Details: This execution includes user-provided project names as command-line arguments, which presents a potential command injection vector if the input is not strictly validated by the agent before shell execution.

generate-clues