Free resource

The Prompt Engineering Guide

Seven things a prompt needs, a five-step play for researching a subject properly, and the point where prompting should stop and automation should take over.

Get the Prompt Engineering Guide

Twenty three pages, three sections, each usable on its own.





One confirmation email first, then the document. Unsubscribe whenever you like.

Pages of the Prompt Engineering Guide fanned out: the cover, an annotated prompt showing all seven elements, and the Find the gap play with a three-circle diagram and the prompt that compares competing results.

The document

Seven things a prompt needs, and the one most people leave out

Most disappointing AI output is a prompt problem rather than a model problem, and the fault is usually one of seven specific omissions. The guide names all seven and shows what each looks like written down.

It also carries a five-step research play and a section on the point where prompting should stop and a rule should take over, which is the part almost no AI guidance covers.

What is inside

Three sections, each usable on its own

You can read any one of these without the others. Which one you need depends on whether the problem is the prompt, the research behind it, or the fact that you are doing the same thing by hand every week.

The anatomy

Seven elements a prompt either has or does not. The ones missing predict exactly how the answer will disappoint you, and each is shown written down rather than described.

Worked examples

Nine before and after pairs across goal, specifics, tone, keywords, structure and context. Everything added is something you already knew and had not said.

The review loop

Five questions to ask a draft before you either use it or start again, and why replying beats restarting.

The research play

Five prompts in sequence that build an outline from what is already ranking rather than from what the model assumes about the subject.

Prompt to workflow

Parameterising a prompt you re-run, chaining, where a prompt sits in a process, and when a model is the wrong tool entirely.

Six templates

Product description, blog post, landing page, email, rewrite and critique. Fill the brackets and all seven elements are already present.

Two pages side by side: before and after rewrites of weak prompts tagged with the elements each adds, and more rewrites with an example of telling the model what already failed.

How it works

Start from what is going wrong

The guide opens on a two-minute diagnostic rather than a contents page, because the four common complaints about AI output have four different causes.

  • Name the complaint

    Bland, off-target, repetitive or confidently wrong. Each one points at a different section.

  • Check the prompt against five questions

    Length, audience, shape, what only you know, and whether you replied to the first answer or started over.

  • Fix the element that is missing

    Not the whole prompt. One of the seven is almost always the thing you were about to complain about.

  • Then decide whether to automate it

    A prompt you have run twice is an asset. A task with one correct answer should not be going through a model at all.

By the numbers

What you are getting

elements, each shown written down
7
step research play with the prompts kept verbatim
5
templates with the variable parts bracketed
6
pages across three sections
23

Before you ask

Questions people send us about this one

Is this tied to one AI tool?

No. Everything in it works the same way in ChatGPT, Claude or Gemini, because it is about what you hand the model rather than which model it is. There is a section on what the newer models changed and what they did not.

I already prompt reasonably well. Is there anything here?

The research play in Section 2 is the part experienced users tend to find new. It is five prompts in sequence rather than one, and the fourth one, which asks specifically what the competition has not answered, is where the value sits.

Does it cover automation?

Section 3 does, and it is deliberately honest about the limits. A working automation is mostly rules with a model at the one or two points that need reading comprehension, and the section says which tasks should never go through a model.

How long does it take to work through?

The diagnostic is two minutes and Section 1 is about ten. The research play takes an afternoon the first time you run it and considerably less after that.

That was the argument. The document is the part you work from.

Printable, sourced line by line, and built to be filled in rather than read. Nobody on our list gets sold to for reading something free.





One confirmation email first, then the document. Unsubscribe whenever you like.