Prompt engineering is the practice of designing, testing and refining the instructions you give an AI model so that it produces reliable, usable output every time, not just once. Where a single prompt is one request, prompt engineering is the process of improving that request until it works across many real cases.
How prompt engineering works
The work looks less like clever wording and more like a small test programme. A typical cycle runs like this:
- Define the job and what good looks like. For example: meta descriptions of 120 to 155 characters, British English, one benefit and one reason to click, no prices unless supplied.
- Collect test cases. Ten or twenty real inputs, including awkward ones: a page with almost no copy, a product with a long technical name, a service with a regulated claim.
- Write a first prompt with context, task, source material, format and constraints.
- Run every test case and score the results against your definition, ideally in a spreadsheet so you can compare versions.
- Change one thing at a time and rerun: add an example, tighten a rule, put the instruction after any long source material rather than before it, split the task into two steps.
Several techniques come up repeatedly. Few-shot prompting shows the model two or three examples of the output you want. Asking for structured output (a table, or labelled fields) makes results easier to check and paste into other tools. Breaking a large task into stages, such as outline first and then each section, usually beats one enormous prompt. Some tasks benefit from asking the model to reason step by step, a technique known as chain of thought. And settings outside the prompt, such as temperature, change how predictable the output is.
Why it matters
Most marketing uses of AI are repetitive: hundreds of product descriptions, a weekly batch of social posts, ad variations for each campaign, summaries of call notes. A prompt that works four times out of five creates a fifth job of finding and fixing the failures, which is often harder than writing from scratch. Engineering the prompt up front is what turns AI from an occasional shortcut into a dependable step in a process.
It also protects you. A tested prompt that says “use only the facts supplied; if a fact is missing, write MISSING” is far less likely to produce invented claims than an improvised one. For UK advertisers, that matters because claims in ads and on websites must be substantiated under the CAP Code, and the business is responsible regardless of which tool drafted the copy.
There is an honest limit to this, too. Prompt engineering cannot give a model knowledge it does not have. If the output keeps getting your product range wrong, the fix is to supply the range, not to find the perfect phrasing.
Common mistakes
- Testing on one example. A prompt tuned on your best-selling product often fails on the other 200.
- Changing several things at once So you never learn which change helped.
- Chasing magic phrases. Copying a “secret” prompt from social media rarely beats a clear brief built around your own material.
- No written definition of success Which turns every review into a matter of taste.
- Forgetting model updates. Providers change models, and a prompt that worked in spring may drift by autumn. Rerun your test cases after a model change.
How to act on it
Pick one repetitive task that costs your team real time each month. Write down, in a few lines, what an acceptable result looks like. Gather ten real inputs and build a simple spreadsheet: input in one column, output from each prompt version in the next, and a pass or fail with a note. Two or three rounds is often enough to get a prompt that passes nine cases out of ten.
Once it works, store the final prompt with its test cases and the date you last checked it. If several people use it, move the standing rules (voice, spelling, banned claims) into a system prompt or a project’s custom instructions so nobody has to paste them each time. Deciding which marketing tasks are worth this effort, and which are better done by hand, is part of the digital marketing strategy and consulting I offer.