The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled Python script using uv run. User-controlled input (componentId) is interpolated into the shell command, which could lead to command injection if the input is not properly sanitized by the agent.
Evidence: uv run {SKILL_ROOT}/scripts/quick_validate.py ./fluxwing/components/${componentId}.uxm {SKILL_ROOT}/schemas/uxm-component.schema.json in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data (component names, properties, and styles) and interpolates them directly into prompts for subagents without using boundary markers or sanitization, creating a surface for indirect instructions to influence subagent behavior.
Ingestion points: Component names, types, properties, and visual styles provided by the user in SKILL.md.
Boundary markers: Absent from the task prompts.
Capability inventory: Write tool for file creation, Bash tool for validation scripts, and Task tool for spawning subagents.
Sanitization: No explicit sanitization or validation of user-provided strings before interpolation is documented.
[DYNAMIC_EXECUTION]: The skill uses the Task tool to dynamically generate and execute logic for subagents based on runtime variables.
Evidence: Multiple Task definitions in SKILL.md that construct prompts using ${componentName}, ${componentType}, and other user-controlled variables.

Fluxwing Component Creator