The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The script scripts/init_skill.py uses chmod(0o755) to grant execution permissions to boilerplate scripts generated during the initialization process. While this is a common developer workflow for creating functional tools, it involves dynamic modification of file system permissions.
[INDIRECT_PROMPT_INJECTION]: The skill implements a scaffolding workflow where user-provided input (such as skill names, descriptions, and interface overrides) is interpolated directly into the contents of generated files (SKILL.md and agents/openai.yaml).
Ingestion points: Untrusted data enters the context via the skill_name argument and --interface key-value pairs in scripts/init_skill.py and scripts/generate_openai_yaml.py.
Boundary markers: The templates use standard [TODO] placeholders but lack explicit security delimiters to prevent embedded instructions in user-supplied strings from being interpreted as logic in the resulting files.
Capability inventory: The skill has the capability to create directories (mkdir), write files (write_text), and modify permissions (chmod) via scripts/init_skill.py.
Sanitization: Skill names are normalized using a regex to lowercase hyphen-case, and interface values are escaped via yaml_quote (escaping backslashes and double quotes), which provides basic structural protection but does not prevent logical instruction injection.
[METADATA_POISONING]: There is a discrepancy in author attribution; the SKILL.md YAML frontmatter claims the author is OpenAI and points to an OpenAI GitHub repository, whereas the skill's operational context identifies the author as JetBrains. This could lead to misjudgment regarding the origin and vetting of the skill's code.
[EXTERNAL_DOWNLOADS]: The skill references resources from OpenAI's official GitHub repository (github.com/openai/skills). This reference is documented neutrally as a link to a well-known service for documentation and source code.

skill-creator