Applied Language Models for AI Builders: Behaviour, Workflows, Agents

A model reasons like a graduate student one moment and cannot count the letters in a word the next. It writes you a working program, then insists that 9.11 is larger than 9.9. It follows a careful instruction perfectly all week, then one Tuesday ignores half of it for reasons nobody can fully explain. You tell a sales bot that aliens are abducting you and it sends you a Calendly link. These are not bugs waiting to be fixed. This is what the thing is, and almost nobody is teaching people how to work with it. What if you understood these models well enough to get what you want from them, predict how they will behave, and see the moment they are about to let you down?

Who is the course for

This is for people who use language models and have begun to feel how little they can predict. You get a good result and cannot say why the next one is worse. You have watched the same request behave differently on different days, or on different models, and had no way to reason about it. You do not need a software background, and you do not need to write code. This course is far more about language and judgment than it is about technology. Technical people are welcome, because understanding how these minds work is not something engineering prepared them for either. What you need is to be working with these models on something real, and to have decided that guessing is no longer good enough.

What we work on

The course lives in the gap between a lucky result and one you can trust. That is where most of the difficulty is. You add more instructions to a prompt and the output gets worse. You phrase the same request two ways and get two different answers. The model states something false with complete confidence, and you cannot tell it apart from the truth. These are common problems, and learning to see them coming is most of what separates someone who can get a clever demo from someone who can be trusted with what they built.

To get there, you give up the idea that you are configuring software. These models behave more like an unfamiliar kind of mind: responsive to tone and framing in ways nobody designed, capable and hopeless in the same breath, and moved almost entirely by words. We treat language as the working instrument it is, the lever you actually have, not a set of tricks to memorise.

We use these models throughout, hands-on from the first session. The aim is not to admire them or fear them, but to understand them well enough to direct them, and to know the moment they are about to go wrong.

Week 1 — Meeting the machine: what it is and how it thinks

Week 2 — Behaviour: why it does what it does

Week 3 — Getting reliable results

Week 4 — Where it breaks, and how it's fooled

Week 5 — Autonomy: letting it act on its own

Week 6 — Bringing it together