Instruction Class (Service)

Author instruction classes — reusable AI capabilities and prompts — that run against a remote LLM endpoint, and register them for use across IVYX Studio.

What you can do

• Configure the remote LLM endpoint: enter its URL and API token, save them, and test the connection.
• Define an instruction class: ID, role, system prompt, a target model (for example gpt-4 or claude-3-opus), temperature, max tokens, and response format (text or JSON), with an optional output schema.
• Edit an existing instruction class — registering an updated one replaces the old definition.
• Compare models: run one instruction class across any set of AI cores and read the answer, duration and eval score side by side per core.

Instruction classes you register here show up in the LLM Capabilities panel under Instruction Classes, where you list and manage them.

Getting started

Open Instruction Class (Service) from the command palette, or from the LLM Capabilities panel's "Add Remote" action under Instruction Classes.

Note

The API token is stored through the host Secret Manager, never in the settings file.

Skills

A skill is an instruction class in mode: instructions. Calling it hands the agent its instructions instead of asking a model, so the agent follows your procedure with its own tools and the conversation it already has. Write one in the editor: set Mode to Hand the agent instructions, give it one sentence of Description (that sentence is how a model decides to use it), and list the files it should carry under Resources. Pick it in the assistant's composer with /, or let the agent choose it from the catalogue. Every call is governed like any other capability: a policy rule can deny it, and the audit trail records which version of the instructions was handed over.