The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/scaffold.py executes the pixi run validate command using the subprocess.run method to verify the repository state after scaffolding operations.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection during the scaffolding process where untrusted user input is written to persistent files.
Ingestion points: The skill collects a description and other metadata from the user which are passed as arguments to scripts/scaffold.py (referenced in SKILL.md).
Boundary markers: The generated SKILL.md uses YAML frontmatter delimiters (---), but the script fails to escape double quotes when interpolating the description field into the YAML block.
Capability inventory: The scaffold.py script has the capability to create new directories, write the SKILL.md file, and append entries to registry.yaml (documented in scripts/scaffold.py).
Sanitization: While the name and tags fields undergo regex validation, the description field is not sanitized or escaped, enabling an attacker to inject new YAML keys or malicious instructions into the created skill context.

sciagent-skill-creator