Most people think AI learns in one place it actually learns in three

Most people think AI learns in one place. It actually learns in three.

When a client asks me to improve their AI product, the first question is usually about the model. Which LLM. What fine-tuning. That's the wrong place to start.

An AI system learns at three levels: the model (the weights), the harness (the instructions, tools, and code wrapped around it), and the context (external knowledge it can access). Most products only touch the first one.

The harness is where the real design work happens. The instructions you write, the constraints you define, the tools you give it. That's the layer you control directly, today, without an ML team.

I've been treating the harness like a product. Iterating on it the way I'd iterate on a user flow. It compounds fast.

If you use Claude, Cursor, or any agentic tool, open the system prompt. That's your harness. Edit it like a design spec and watch what changes.