Now say we want to simulate a branding expert. The model could draw on data about brand building, reputation management and positioning strategy, and give more precise, detailed advice than a generic system would. If we wanted the LLM to "play" a marketer in a specific niche, it would be worth a great deal to companies working in specialized markets. Feed it specific context through company documents and it gets more precise still. For example, an LLM focused on B2B marketing would suggest different strategies than one specialized in B2C marketing, reflecting the different needs of these two segments.

Picture an LLM playing a digital marketing expert. The model is trained on data from every industry, so it also holds plenty of digital marketing material pulled from books, blog posts and videos. Ask it to play that specific professional and it will "filter its knowledge", giving you advice on pay-per-click advertising strategy, search engine optimization and social media metrics analysis without drifting into broad generalizations.

Now say we want to simulate a branding expert. The model could draw on data about brand building, reputation management and positioning strategy, and give more precise, detailed advice than a generic system would. If we wanted the LLM to "play" a marketer in a specific niche, it would be worth a great deal to companies working in specialized markets. And if you give the model specific context through company documents, its advice gets even more precise. For example, an LLM focused on B2B marketing would suggest different strategies than one specialized in B2C marketing, reflecting the different needs of these two segments.

Role-play gives LLMs a clear frame of reference. Instead of searching a huge pool of data for every prompt, the model uses its assigned role as a guide, a lens through which it reads your request and answers it.

Role-play gets especially useful once you move into an agentic setup, where each agent has a well-defined role and together they form a team that works in sync. Here, role-play techniques let us define each job, and each team member's point of view, with precision.

Getting answers closer to what you expect.

An LLM that 'thinks' like a content marketing expert might, for example, suggest content that resonates with a specific audience, or the best ways to distribute it. And you can go further than defining an area of expertise.

Personality

Give the AI a 'personality'. In marketing, that could mean deciding whether the LLM should sound like a seasoned industry pro or maybe a young innovator who keeps up with every new trend. This 'personality' shapes the tone, style and depth of its answers.

Relationships

Give it clear context, such as the kind of company it identifies with, the marketing goals it's working toward or the quirks of the market it operates in, and the model's answers get more precise.

For example, if the model plays a strategist for a luxury brand putting together a newsletter, that context will guide its recommendations.

{
    "ruolo": "Digital Strategist",
    "personalità": {
        "tipo": "professionista sofisticato",
        "attributi": {
           "tono": "elegante",
            "stile": "formale",
            "profondità": "dettagliata"
        }
    },
    "contesto": "Marketing Digitale per un Marchio di Lusso nel 2024",
    "obiettivo": "Fornire informazioni sulle ultime tendenze e strategie nel settore del lusso."
}

What follows is a preview of my guide "Anatomy of a Prompt", where I share my personal method for structuring a prompt. I built it over 2 years of testing, thousands of tests in all, plus endless hours reading research papers.

Besides the "role" or "character" the model takes on in the conversation, which I covered above, there's plenty of other information you can put in a "system prompt" to shape the answer. Here are a few:

1. Format: say whether you want a short answer, a detailed one, a list, paragraphs and so on.
2. Tone of voice: say whether you want the answer formal, informal, friendly, professional, enthusiastic, neutral and so on.
3. Target audience: this helps the model pitch the answer to whoever will read or receive it (e.g. teenagers, professionals, parents and so on).
4. Level of detail: you can ask for a general overview, an in-depth answer or something in between.
5. Examples or analogies: if you want the answer to include concrete examples or analogies that make it easier to follow.
6. References or sources: if you want the model to cite established sources or references to back up the answer.
7. Geolocation: name a specific region, country or city if you want geographically relevant information.
8. Exclusions: any topic or detail you'd rather the answer leave out.
9. Comparisons: if you want to compare two concepts, products, strategies and so on.

The problem is that today's models struggle to "read" and understand long prompts, so you need to give them structure. But that's a guide for another day

If you found this useful, help us grow. The goal is to become a big community, hungry for knowledge, so we can ask companies for discounts, early access, event invitations, funding to keep the project going and perks for the members of Marketing Hackers, which will always be free.

The more of us there are, the more we'll get. So if you like the project, share it with your colleagues and friends and invite them to sign up, so the knowledge spreads and nobody gets left behind.

Stay Tuned!