How To Design Intelligent Systems

This workshop treats an intelligent system as an arrangement of parts, each doing one kind of work: some steps are rules written in code, some are judgements only a person can make, and some are small models that read and choose but do not decide.

The method starts from a task you already do. You break it into steps, say for each step who or what should do it and why, and write down what is lost when a whole task is split into parts. Where a model is used, its job is kept narrow: it answers from a fixed set of choices, or says it does not know, and what follows is set by code and by people.

A person's judgement is what the system learns from. You define what each answer means, label the examples yourself, and test what you build on examples it has not seen, against the simplest rule that could do the same job. Nothing counts as working until its errors are recorded and explained. What you leave with is a system in which every decision can be traced to a rule someone wrote or a judgement someone made.

Online, every Saturday for four months, from 5 December. Free, as part of Koher's outreach. You need a coding agent set up on your own computer, such as Claude Code, OpenAI Codex or Gemini CLI.

Students from any university or college, at any stage of their studies, and working professionals can apply by Tuesday 1 December. There are 12 places. I read every application before accepting anyone.

Closes 1 Dec 2026, 23:59

I will write to you here when I accept

For example: B.Des, Interaction Design; or your organisation and role

Any semester or stage, or a working professional

One or two lines. It can change in the workshop.

Any one, free or paid: for example Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot, Cursor, Windsurf, opencode, Qwen Code, Cline or Aider

checking you are a person…