The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches technical documentation from a well-known service at https://claude.com/docs/llms-full.txt to align its guidance with platform-specific constraints. This is a legitimate use of official resources for context loading.
[INDIRECT_PROMPT_INJECTION]: The skill describes the implementation of protocol primitives like 'Elicitation' and 'Resources' (URI templates), which ingest data from user forms and workspace files. These represent an inherent attack surface for indirect prompt injection in the developed servers.
Ingestion points: references/resources-and-prompts.md (reading files via URI templates) and references/elicitation.md (collecting structured user input via forms).
Boundary markers: The provided code snippets are functional scaffolds and do not include specific instruction delimiters or 'ignore' warnings for the data they process.
Capability inventory: The skill outlines capabilities for tool execution, local filesystem access, and LLM sampling via the host application.
Sanitization: While the tool-design.md reference provides advice on writing tight schemas and descriptions to prevent misuse, the code scaffolds themselves do not implement explicit input sanitization logic.

build-mcp-server