The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by reading local project files. Ingestion points: Workflow Step 1 in SKILL.md reads local files like package.json and README.md. Boundary markers: The skill does not use technical boundary markers for ingested file content. Capability inventory: Workflow Step 3 in SKILL.md allows writing to the local file system (docs/scratchpad/CLAUDE.md). Sanitization: The risk is mitigated by an interview process that acts as a human-in-the-loop review of the content before any file is written.

scaffold-claude