Introduction

I have been building with large language models for long enough to have collected a pile of habits. Some of them are obvious in hindsight. Most of them cost me a day or two to learn. This book is where I write them down.

It is not a tutorial and it is not a survey. There are better versions of both elsewhere, written by people with more time. What is harder to find is the middle layer: the small, unglamorous techniques that separate a model that almost works from one that does the job every time.

Who this is for#

Someone who has already got past the first conversation. You know how to call an API or drive an agent. What you want now is the working knowledge – how to structure a prompt so it survives contact with real input, when to reach for tools instead of more instructions, what to do when the output is nearly right.

If you have never used a model before, start elsewhere and come back. Nothing here assumes deep expertise, but it does assume you have felt the frustration these techniques are meant to solve.

The one rule#

Everything in this book is downstream of a single idea:

A language model is not a program you instruct. It is a system you steer.

Programs fail loudly and in the same place every time. Models fail quietly, in different places, and often in ways that look like success. So the techniques that matter are not the ones that make a model clever. They are the ones that make its failures visible – and then cheap to correct.

That is why so much of what follows is about structure rather than wording: where to put the constraint, how to shape the output so a mistake shows up immediately, when to stop asking and start checking.

How to read it#

Chapters are short and self-contained. Each one covers one technique, why it works, and where it stops working – because the boundary is usually the useful part.

Read it in order if you want the argument to build. Use the contents or the search box if you came looking for something specific. Nothing later depends on anything earlier, so skipping is safe.

A note on honesty#

Where a technique is a rule of thumb, I say so. Where I am guessing, I say that too. Models change quickly and some of this will age badly – every chapter carries the date it was last touched, so you can judge for yourself how much salt to apply.

If you want to know how the book itself is put together, the next chapter covers it: how this book is written.