Prompt Engineering

What Prompt Engineering is

Prompt engineering is the practice of structuring a prompt deliberately, instead of typing the first phrasing that comes to mind, so that the result is closer to what is needed. The word “engineering” can mislead. It does not mean code, and it does not need a technical background. It means treating the instruction as something with parts that can be checked and improved, in place of a sentence that either works or does not.

The site teaches it as a practical skill in the foundational courses, not as theory. Participants write their own prompts on their own tasks, send them into the chat and improve them live. That is the point of the word “practical”: the skill is in doing it repeatedly on real work, which is why the exercise is your own material and not a worked example from somebody else’s business.

The method: four parts

The prompting workshop starts from one reliable method, the four essential parts of a prompt: goal, context, source and expectations. The HR course puts the same four as questions a person can ask before typing: what is wanted, who it is for, based on what, and in what shape.

The goal says what the result is for. The context supplies the background and the reader. The source names the material the answer should draw on, and the expectations describe what good looks like, including format, length, tone and what to avoid. A prompt with only a goal gets a generic answer, because the tool has filled in the rest with the statistically likely choice.

The brief article on this site extends the same idea to six things a good brief covers: the job the tool is doing right now, the deliverable and its reader, the background a newcomer would need, what to avoid, one example of a result you have accepted before, and how the output will be judged. The two lists describe one habit. Both say the request is a small part of what the tool needs to know.

Prompt Engineering and Context Engineering

Bad AI output is rarely a wording problem. It is usually a missing-context problem, and our article on briefing an AI tool like a new hire sets out why. That observation has picked up its own name, context engineering, which sits alongside prompt engineering and does not replace it.

The two fit together. Prompt engineering is still real and still useful, in the way that phrasing an instruction to a new hire clearly still matters. A muddled sentence will produce a muddled answer whatever surrounds it. But a beautifully worded instruction with nothing else attached still leaves the tool guessing at the parts nobody said out loud, and a plainly worded one with the right material attached usually does not. If you have time to improve only one thing, improve the context.

Techniques and what each is for

Once the four parts are in place, the workshop teaches a menu of techniques. Think of them by what each one does.

Some change who is answering. Role prompting and personas give the tool a job or a style to answer in, and the Copilot Chat course teaches role and context prompting, giving the assistant a backstory. Some show the tool what you want. One-shot and few-shot prompting supply one or several examples, which is often quicker than describing the result. Output format control fixes the shape of the reply, and asking for sources and citations makes the answer easier to check.

Some reverse the direction. In question and answer prompting the tool interviews you before it answers, so it collects the context you left out. Assumption handling makes the tool state what it is assuming. A grounded version puts the questions to a document you attach, with a fallback so that when the document is silent the tool says the answer is not specified in the document. That is grounding at its smallest scale.

Some change how the tool reasons. Chain of thought asks for the working in steps, and tree of thought explores several branches before choosing. Prompt chaining splits a large task into a sequence in which each output feeds the next, and the workshop’s example runs a meeting transcript, an agenda and a template through a seven-step chain. Meta prompting goes one step further and uses the AI to write and improve your prompts.

No technique replaces the four parts. They work on top of them.

A worked example

Start from the bare instruction “Write the client update.” The tool has a goal and nothing else, so it guesses the reader, the length and the tone. Add the four parts and the guess disappears. The goal is a short update on this month’s progress. The context is who reads it and what last month’s version looked like. The source is the meeting notes you attach. The expectations are the length and tone, plus a request to flag anything that slipped.

Then add techniques only where the result still falls short. If the tone is wrong, give the tool a role or attach one example of an update you have accepted before. If the notes are patchy, ask it to list what it is assuming or to interview you before it writes. If the update needs facts from the notes, tell it to answer only from them and to say when the notes do not cover something. Each addition fixes a named gap, and none of them is a cleverer sentence.

Common failures and what they look like

The brief article describes a recognisable pattern when one part goes missing. A role with no goal produces confident work aimed at the wrong reader. A goal with no constraints produces something usable that still needs a tone pass. Constraints with no example produce something that obeys every rule and still does not read right, because “no jargon” and “here is what it should sound like” are different instructions.

Reading a failed result against that list is quicker than rewording at random. It also explains why two people can type the same clear sentence into the same tool and get very different answers. One of them has built the surrounding habits, and the other is starting from a blank line every time.

Iterating and debugging

A first prompt is a draft. The workshop teaches prompt debugging, why prompts fail and five questions that fix them, so that a poor result becomes a diagnosis instead of a guess at new wording. It is a faster route than starting again, because most failures have a cause that can be named: missing context, a vague goal, no source, or no statement of what good looks like.

The assistant can take part in that work. The Copilot Chat course teaches using the assistant itself as a prompt coach, and the editors’ workshop builds a prompt improving agent. The pattern is the same as good editing. Write the draft, ask what is missing, add it, and compare the results.

From a Good Prompt to a Reusable Tool

A prompt that works once is worth keeping. The move from writing prompts to reusing reliable ones is the subject of the prompt library: a kept set of prompts already shown to work, reused rather than rewritten. A prompt template is a fixed structure with gaps to fill in. The ChatGPT Enterprise course turns prompts that work into custom GPTs, an AI prompter among them.

Some prompts then stop being prompts. A scheduled prompt runs on its own at set times, such as a daily digest of your emails, meetings and files, or a weekly search on a topic you follow. An AI agent holds its own instructions and takes several steps. The workshop closes on a decision framework for choosing between a prompt, a template, a scheduled prompt and an agent, because the answer depends on how often the task repeats and how much judgement it needs. Standing rules belong in custom instructions, which carry the context that never changes so each prompt only has to supply what is new.

What Prompt Engineering cannot do

A model predicts the most likely next token. It has no step that checks a claim against a record, so a wrong statement and a right one come out of the same process. A better prompt narrows the room for an invented answer, because a real example, a named reader and a clear standard give the tool something concrete to match. It cannot supply a source the tool does not have. That failure is hallucination, and the answer still needs a human to check it. The courses treat humans in the loop as part of the method, and the HR course teaches how to make a tool quote its source and flag what is missing.

What goes into a prompt matters as much as how it is worded. Everything you type is data. Our article on why banning AI at work backfires argues for a short rule drawn by kind of data, and for an approved tool good enough that staff do not reach for a personal account. A well-engineered prompt that carries a client contract into an unapproved tool is still a mistake.

Where it is taught and where it shows up in an Audit

Prompt engineering runs through the core courses. The Art of Prompting workshop, over two half days, goes deepest, and it is not tied to one tool. The Copilot Chat day and the ChatGPT Enterprise course teach it on the tool participants already hold, and the editors and authors workshop applies it to writing and editing against a house style. The set courses list the full catalogue.

An audit meets it from the other side. Repeated prompts are a signal. When a team keeps rewriting the same prompt, the task behind it is a candidate for a template, a scheduled prompt or an agent, and that is the kind of workflow an audit identifies. Uneven prompting skill across a team is another signal, because one or two people who write good prompts and everyone else guessing tends to show up first in whichever process is examined closely.