The Agent Skills Directory

[PROMPT_INJECTION]: The skill functions as a generator that processes user-supplied requirements to create instructions and metadata for new skills, which constitutes a surface for indirect prompt injection.
Ingestion points: User input regarding skill purpose, activation cues, and expected outputs gathered during the scaffolding process in SKILL.md and agents/scoper.md.
Boundary markers: The skill uses a structured template for the generated SKILL.md file but does not implement specific delimiters or escaping for user-supplied strings within the generated instructions.
Capability inventory: The agent is empowered to write files to the repository and execute validation commands (such as linters or catalog-specific scripts) mentioned in SKILL.md Step 6.
Sanitization: The skill includes explicit validation rules for metadata fields, such as enforcing kebab-case for names and character limits for descriptions, which serves as a basic sanitization layer.
[SAFE]: No malicious patterns or suspicious behaviors were detected in the skill's own code or instructions. It follows established best practices for skill authoring.
[REMOTE_CODE_EXECUTION]: The instructions explicitly forbid the use of remote-code execution patterns, specifically warning against scripts that use 'download-and-run' methods.
[DATA_EXFILTRATION]: The skill includes a 'Safety & constraints' section that explicitly prohibits hidden network access and the harvesting of secrets or credentials.

fusion-skill-authoring