The Agent Skills Directory

[SAFE]: The skill implements a secure authoring workflow that emphasizes safety best practices. For example, references/iteration-evidence.md explicitly instructs to redact secrets and personal data from examples, and references/claude-code/hook-backed-skills.md provides warnings about shell execution risks and permissions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from external sources such as documentation, git history, and issue logs to synthesize new skill instructions. This is a functional requirement of a meta-authoring tool but represents an attack surface for indirect prompt injection.
Ingestion points: External source materials including domain documentation, upstream implementations, and repository history as defined in references/synthesis-path.md and references/source-discovery.md.
Boundary markers: The instructions focus on imperative voice and structured output but do not specify mandatory delimiters for ingested content beyond anonymization.
Capability inventory: The skill performs file system writes to create skill directories and executes a local validation script (scripts/quick_validate.py).
Sanitization: references/iteration-evidence.md requires anonymizing examples and redacting sensitive information like credentials and private URLs.

skill-writer