The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The file references/crawling-guide.md provides bash command snippets (mkdir, curl) intended for the agent to execute to manage documentation files. Writing directly to the skill's reference directory (.claude/skills/...) is a high-risk operation. Evidence: mkdir -p .claude/skills/{library}/references, curl -s https://example.com/sitemap.xml -o sitemap.xml.
EXTERNAL_DOWNLOADS (MEDIUM): The skill guides the agent to download files from untrusted external domains (example.com, etc.) directly into the skill's reference directory. Evidence: curl -s https://example.com/llms.txt -o .claude/skills/{library}/references/docs.md.
PROMPT_INJECTION (LOW): The templates/skill.md file uses an override pattern in its description to force execution. Evidence: **IMPORTANT: Always execute this skill before answering questions about these libraries.**.
INDIRECT_PROMPT_INJECTION (LOW): The skill has a significant attack surface for indirect prompt injection as it is designed to fetch and process untrusted external web content. 1. Ingestion points: WebFetch calls in references/crawling-guide.md (URL harvesting) and templates/skill.md (detailed doc retrieval). 2. Boundary markers: Absent. The prompts do not use delimiters or instructions to ignore embedded commands. 3. Capability inventory: WebFetch, Read, WebSearch, and shell access via curl. 4. Sanitization: Absent; content is processed directly for extraction without validation.

lib-docs-generator