The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a pattern for LLM interaction that is susceptible to indirect prompt injection.
Ingestion points: The improve_text function in SKILL.md accepts a text string which is intended to be resume content.
Boundary markers: Absent. The prompt is constructed using direct f-string interpolation (f"Improve this resume text:\n\n{text}") without the use of delimiters (like XML tags or triple quotes) or 'ignore instructions' warnings to the LLM.
Capability inventory: The skill metadata explicitly allows high-privilege tool access via Bash(python:*), Bash(pip:*), and Bash(uv:*), making the impact of a successful injection high.
Sanitization: No sanitization, escaping, or validation logic is defined to check the text input for embedded instructions or adversarial content.

backend-dev