The AI Practice Library for Marketing and Communications

• One hundred fill-in prompts in eight categories, running from the brand voice foundation and social content through to website, advertising, brand monitoring and AI policy
• Category 1, the brand voice foundation: twelve steps that together produce one briefing, plus the blind test that makes it pass or fail
• Category 8, AI policy: the texts with which you set out, after the fact, what may and may not go into a prompt window
• Framework chapters on instruction, on release and on the legal framework, including where the labelling obligation begins and where it stops
• A separate chapter that is literally called the position of the editorial board, so that you can see where the sources stop and where a view begins
• Every prompt in the same five parts: the purpose, when to reach for it, the full prompt in a copyable block, a note for each placeholder, and the release check tailored to that specific prompt
• Every prompt built from the same six blocks: Role, Voice, Context, Task, Format and Release
• Five diagrams showing the structure, the sequence and the release gates at a glance
• Fifty-two verified sources across nine domains, approximately 1,400 references in the text, research cut-off date 4 August 2026

The book sets out a scene before the hundred prompts begin, and you have probably played it yourself. A fictional installation company with twelve engineers introduces a maintenance subscription and wants to put something about it on LinkedIn. The instruction to the model reads: "Write a LinkedIn post about our new maintenance subscription. Make it catchy and professional." You know what comes out before you read it. An opening line with a rhetorical question, three benefits each preceded by a tick, an invitation to comment, and somewhere the words "complete peace of mind".

The book then does something uncomfortable. It goes back over the instruction, word by word. "Catchy" and "professional" are not instructions but wishes, and every company in the Netherlands would endorse them. It did not say who reads this. It did not say what the subscription actually covers. It did not say why the company offers it, and therefore did not say that these customers buy on reliability and that enthusiasm works against you with them. There was no length, no structure, no control. You did not measure what the tool can do. You measured what happens when a tool knows nothing about you.

The comparison the book uses for this comes from the documentation of one of the model makers itself: treat the model as an exceptionally capable but brand new colleague who does not know your standards or your working methods. That colleague has not read your brand guide. He does not know that this company writes "service visit" and never "service intervention", or that it writes "what it costs" and never "investment", and he does not know that the previous campaign came apart on precisely that one word. He fills the empty spaces with the average of everything that exists on your subject, and the average of everything sounds like nobody.

That is where the story turns. The hundred prompts in this library do nothing magical. They supply the briefing you never gave, every time, in the same order, so that you do not have to remember it. Every prompt consists of six fixed blocks: Role, Voice, Context, Task, Format and Release. Role and Voice determine how it sounds. Context and Task determine what comes out. Format and Release determine whether you dare to send it.

And then the second mistake, the one from the hook. There is no general obligation to mark all AI content. In its guidelines the European Commission expressly names AI-edited text that forms part of an advertisement or a product description as an example falling outside the scope of that provision, with claims about health, consumer safety and sustainability among the exceptions. For image and video it is different, because there you may genuinely be dealing with a deepfake. The book works out where that line runs, so that you stop labelling where you do not have to and start looking where it does matter.

And that looking is the second half of the book, the half that is less visible and weighs more heavily over time. Good output is worthless as long as you do not dare to publish it. The real brake in practice is not the quality of the text but the doubt afterwards. Is that figure the model mentioned correct. May this statement appear on our site in this form. Have I just pasted customer data into a tool I do not control. That is why every prompt in this book ends not with the text but with a release: four gates the model runs across its own work, and one question you answer afterwards. That question cannot be delegated. Do I put my name to this.

• Why the building block that carries your brand is missing from almost every existing prompt model, or sinks to the level of a parameter somewhere between length and format, and what changes the moment it is given a place of its own.
• How twelve steps together produce a single brand voice briefing that is concrete enough for a new colleague to write from without speaking to you, and which you paste into every subsequent prompt from that moment on.
• The blind test that makes your briefing pass or fail on a text that has nothing to do with your field, for example an explanation of how to separate waste, and why every open decision the model reports there is a rule you have yet to write.
• Why "no jargon" does not work and one sentence of explanation behind it does, because a prohibition covers one case and an explained reason also covers the cases you forgot to write down.
• Part 4 of the avoid-list, the list of words and constructions the model must not use, the part almost everyone skips, and why without it the model deletes precisely the words that make your brand recognisable.
• When you do have to label and, more important for your working day, when you do not, and why a talking mouse in a cheese advertisement falls outside the deepfake definition while an AI-generated image of your product on the packaging may well fall within it.
• The four gates every prompt closes with, and the two reported data breaches that show why the third gate is about what you enter and not about what you publish.
• One hundred fill-in prompts in eight categories, plus the framework chapters on instruction, release and the legal framework, at approximately 96,000 words.
• Five diagrams that show the structure, the sequence and the release gates at a glance.
• Both language editions, Dutch and English, complete and both included, in PDF and Word, so that you can copy, amend and reuse the prompts and your brand voice briefing directly in your own documents.

This book is written for three kinds of reader, and what they share is that they are themselves the sender of whatever goes out.

The self-employed content marketer and the communications adviser in a small agency. You are writer, strategist and editor in one person. There is no team reading along and no lawyer to glance over it. You need time, and you do not dare to take the time you would save as long as you do not trust the output. This book gives you both halves: instruction that sounds like the client you are writing for, and a check you can run without legal training.

The in-house marketing or communications manager at an SME, a municipality, a healthcare organisation or a non-profit. You were handed the AI policy internally without asking for it. Colleagues have been using the tools for a long time, but nobody has written down what may and may not go into a prompt window. You are not the exception there. In research among Dutch marketing professionals, 54 per cent of organisations had an AI policy in 2025, while 29 per cent had guidelines covering both internal and external use. Category 8 gives you the texts with which you make those agreements after the fact.

The entrepreneur who does their own marketing. No department, no agency, no copywriter. What you publish is directly what your brand is, and your name is underneath it, so you are the one who has to be able to answer if somebody asks. For you the release check is not a procedure but the only second reader available.

Where you probably are: you are not a beginner. You tried AI, you got brandless output back, and you concluded that it does not work for your brand. This book assumes the fault lay not with the tool but with the instruction, and works that out across one hundred concrete cases. This edition is less suitable if you are looking for a list of ready-made sentences or one prompt that solves everything. The prompts work without adaptation and work better with adaptation, and that adaptation begins with sitting down properly once for your brand voice.

Every factual statement in the framework chapters carries a source reference. Those references point to a closed set of fifty-two verified sources across nine domains, from the EU AI Act and European supervision to the Nederlandse Reclame Code (Dutch Advertising Code), sector research and prompt methodology. All public, all retrieved again rather than quoted from memory, all open to your own inspection. Taken together, there are approximately 1,400 references in the text.

A separate verification round then set every citation against the source text. Where a statement was too strong, it was corrected. Where it did not survive verification, it was deleted rather than softened. That also produced a list of hard limits the writers had to observe, for example that as at the research cut-off date there is no enforcement practice under Article 50 and that no fines or regulatory rulings under that provision may therefore be cited.

Equally important is what remains visibly separated. Fact stands as fact, with a source attached. Craft stands as craft: the prompts themselves carry no source, because a formulation either works or does not work and no literature exists that can substantiate it sentence by sentence. Opinion stands as opinion, in a separate chapter that is literally called the position of the editorial board, so that you can see where the sources stop and where a view begins.

What you do not get here matters just as much. No percentage uplift, because no research exists that expresses the effect of roles, examples or tone-of-voice instructions on brand consistency in figures. The one measured number that came up in this research comes from a benchmark for programming tasks and says nothing about copywriting. Anyone who does give you that figure has invented it.

Where guidance is not binding, the book says so. Where two studies have measured different populations, their figures stand side by side as two separate observations and not as confirmation of one another. And the book states what it is not. The guidelines of the European Commission are not binding, and an authoritative interpretation of the EU AI Act can ultimately come only from the Court of Justice of the European Union. The book is not legal advice on your own file. The research cut-off date is 4 August 2026, and what changes after that is not in it. Those three sentences are not there out of caution. They are there because a framework that conceals its own limits will let you down at the wrong moment. All examples are generic or expressly fictional. There are no customer stories that do not exist, and the content is attributed to the publisher rather than to individuals.

Fifty-two verified sources across nine domains, from the EU AI Act and European supervision to the Nederlandse Reclame Code (Dutch Advertising Code), sector research and prompt methodology. Approximately 1,400 references in the text. Research cut-off date of the study: 4 August 2026. Every citation was set against its source text in a separate verification round, and statements that did not survive it were deleted rather than softened. Where guidance is not binding, the book says so: the guidelines of the European Commission are not binding, and an authoritative interpretation of the EU AI Act can come only from the Court of Justice of the European Union. This edition is not legal advice on your own file or your own campaign. All examples are generic or expressly fictional.