Cortex · Prompt and context

Output schema in the prompt (`output.doc`)

Signal sends the output schema one of three ways, by what the provider can do: with the request as a JSON Schema (native), as the parameters of a respond tool, or, under emit, as a system message, OUTPUT SCHEMA, after the tool guides. output.doc puts a string of yours in place of that message, or leaves it out with 'off'.

Agent

const Order = z.object({ night: z.string().describe('YYYY-MM-DD'), seats: z.number().int() });

const clerk = defineAgent({ id: 'clerk', instructions: 'Read the order.', output: { schema: Order } });
const ownDoc = defineAgent({ id: 'clerk', instructions: 'Read the order.', output: { schema: Order, doc: '…' } });

// each previewed on three providers: one that takes a JSON Schema with a request,
// one that keeps tool arguments to their schema, and one that mangles them
await clerk.preview('Two seats on Saturday.', { llm });

System messages, by provider

{
  takes a JSON Schema: {
    strategy: 'native',
    system: [
      'Read the order.',
      'FINISH PROTOCOL: To finish, reply with no tool calls…',
    ],
  },
  keeps tool arguments: {
    strategy: 'respond',
    system: [
      'Read the order.',
      'FINISH PROTOCOL: To finish, call the `respond` tool …',
    ],
  },
  mangles tool arguments: {
    strategy: 'emit',
    system: [
      'Read the order.',
      'OUTPUT SCHEMA — your output MUST validate against th…',
      'FINISH PROTOCOL: Your ENTIRE final message must be O…',
    ],
  },
  mangles them, with doc: {
    strategy: 'emit',
    system: [
      'Read the order.',
      'OUTPUT SCHEMA: an order, the night as YYYY-MM-DD.',
      'FINISH PROTOCOL: Your ENTIRE final message must be O…',
    ],
  },
}

Output matches the expected value

Runs in this page on the installed package. The input and the expected output are defined by the example.

Output schema in the prompt (output.doc)

Output schema in the prompt (`output.doc`)