Functor example

functor-union.ts

const selectTopic = functor({
  prompt: ({ query, topics }: { query: string; topics: string }) => `
    User query: ${query}

    Existing topics:
    ${topics}

    Match an existing topic if one fits, otherwise create a new one.
  `,
  returns: [
    {
      type: 'choose_existing',
      description: 'Pick an existing topic that matches',
      schema: z.object({ topicId: z.string() }),
    },
    {
      type: 'create_new',
      description: 'Create a new topic when none match',
      schema: z.object({
        name: z.string(),
        summary: z.string(),
      }),
    },
  ],
});

functor.ts

const extractPatient = functor({
  prompt: ({ text }: { text: string }) => `
    Extract patient info from the text below.
    Return whatever fields are present.

    ${text}
  `,
  returns: {
    description: 'Extract basic patient demographics if present',
    schema: z.object({
      name: z.string().optional(),
      birthDate: z.string().optional(),
      gender: z.string().optional(),
    }),
  },
});

const patient = await extractPatient({ text: documentText });

A small shift that’s changed how we build at Triangle: treat LLMs as functions, not chatbots or agents.

One call. Typed input. Typed output. No conversation, no loop.

We call them Functors.

The schema is the return type. The prompt is the implementation.

What followed was realizing how much of a codebase can be written this way.

A lot of regexes and hand-rolled parsers were just deterministic answers to non-deterministic questions.

Prompts are just another programming language now.

You write functions in prose, then compose them with the rest of your code like anything else: called from a worker, an API route, or a UI handler.

Functors have a stable interface. They behave like any code: testable, versioned.

Need more than one return shape? The prompt stays a typed function, returns becomes an array, and you get a typed discriminated union. Dispatched in prose.

The trick that makes it work: tool calling, used to select and enforce return types.

Each return is a schema plus a short description.

The model thinks it’s choosing an action. We use it to select the shape of the return value.

https://x.com/ninjudd/status/2044098508920369535