The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (LOW): The skill instructs the installation of the raison npm package. Although this is the primary purpose of the skill, it involves an external dependency from a source outside the trusted list.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill defines a workflow where untrusted data (variables) is interpolated into LLM prompts. By design, the SDK disables escaping (noEscape: true), which is common for AI prompts but creates a surface for schema confusion or injection.
Ingestion points: raison.render(id, variables) accepts arbitrary data objects from the host application.
Boundary markers: The provided examples do not demonstrate the use of delimiters (like XML tags or brackets) to isolate user-provided variables within the prompt templates.
Capability inventory: The rendered output is explicitly intended to be used as system instructions for powerful LLMs such as Claude and GPT-4o, as shown in references/examples.md.
Sanitization: Documentation in references/sdk-reference.md confirms that HTML/special character escaping is disabled by default to maintain prompt integrity for LLMs.
[DYNAMIC_EXECUTION] (SAFE): The SDK performs runtime compilation of Handlebars templates. This is the core functionality of the tool and is treated as safe in the context of a prompt management SDK.

raison-sdk