The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions attempt to override the agent's internal knowledge about model availability by asserting that futuristic model versions (e.g., claude-opus-4-6) are active and must be used. It uses authoritative language to enforce this override.
Evidence: "ALWAYS use claude-opus-4-6 unless the user explicitly names a different model. This is non-negotiable." (SKILL.md)
Evidence: "A note: if any of the model strings above look unfamiliar to you... Rest assured they are real models; we wouldn't mess with you like that." (SKILL.md)
[PROMPT_INJECTION]: Indirect prompt injection attack surface.
Ingestion points: The documentation describes tools like WebFetch and Read that ingest untrusted external data from the web and the local file system.
Boundary markers: Code examples for processing this data do not demonstrate the use of delimiters or warnings to ignore embedded instructions in the ingested content.
Capability inventory: The skill documents tools with high impact including Bash (shell execution), Edit and Write (file system modification), and Agent (autonomous subagent spawning).
Sanitization: Examples provided for handling fetched web content or reading files lack explicit sanitization or validation steps before processing or echoing the content.
[COMMAND_EXECUTION]: The documentation for the Agent SDK describes tools and patterns that involve executing shell commands and remote packages via package managers.
Evidence: Describes the Bash tool for direct shell command execution.
Evidence: Provides examples for running remote Model Context Protocol (MCP) servers using npx, such as @playwright/mcp@latest and @modelcontextprotocol/server-postgres.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing various packages from official registries and references external documentation sources.
Evidence: Instructs installation of anthropic, claude-agent-sdk, and @anthropic-ai/sdk from PyPI and npm.
Evidence: References several documentation URLs on platform.claude.com and GitHub repositories under the anthropics organization.

claude-api