There is an enormous difference between using AI for marketing and using it well. Most teams that say "AI doesn’t work for them" are using prompts along the lines of "write a post about our product". Those who get consistent results have learnt to communicate with the models in a way that takes advantage of their architecture.

Prompt engineering is not a technical skill. It is precise communication. And in marketing, precise communication is what we do every day.

The five techniques with the greatest impact

1. Role + Context + Task (the basic triptych)

The most effective prompt for the majority of marketing tasks has three components: the role the model should adopt, the business context, and the specific task.

Mediocre prompt: "Write a follow-up email for a lead who hasn’t replied."

Effective prompt: "You are the commercial director of BAI Marketing, an AI marketing agency in Bilbao specialising in industrial SMEs. The lead is Pablo Martín, director of Automoción Martín in Vitoria-Gasteiz, who visited our voice agent landing page 8 days ago but has not replied to the first email. His website has a load speed of 6 seconds and has no claimed Google Business profile. Write a follow-up email of no more than 120 words that specifically mentions the Google Business issue as a hook, proposes a 15-minute call, and has a subject line that does not begin with ‘Follow-up on’."

The difference in output quality is substantial. The second prompt produces an email you can use directly. The first produces an email that needs rewriting.

2. Few-shot examples for brand voice

Language models learn from the context of the conversation. If you show them three examples of how your brand writes before asking them to write something, the output respects the brand voice far more than if you simply describe it in words.

"Here are three examples of how BAI Marketing writes its prospecting emails: [example 1] [example 2] [example 3]. Write an email in the same tone for the company Fontanería Bilbao S.L."

This is especially useful when the brand voice has characteristics that are difficult to describe but easy to demonstrate: the rhythm of sentences, the level of formality, the use of humour, the tendency to use rhetorical questions.

3. Chain-of-thought for complex analysis

For analysis tasks (interpreting campaign data, evaluating a strategy proposal, diagnosing why a campaign is not working), asking the model to show its reasoning step by step improves the quality of the analysis.

"Analyse this Google Ads campaign data step by step: first identify which metrics are outside the benchmark, then explain what may be causing each anomaly, and finally propose the three priority actions ranked by expected impact."

The “step by step” activates a different reasoning mode in the model that produces more rigorous and less superficial analysis.

4. Output structure for consistency

If the model knows exactly what format the output should take, it produces more consistent outputs that are easier to process (especially useful for automations where the output is used as the input of another process).

"Respond in JSON format with this structure: {subject: string, body: string, cta: string, tone: string, main_keywords: [array]}. Do not include any text outside the JSON."

For teams that want to use AI systematically, building a prompt library with standardised output structures is one of the highest-return investments: anyone on the team can use the prompt and get an output in the expected format.

5. Negative constraints to avoid common mistakes

Telling the model what it should NOT do is as important as telling it what it should do. Models have training biases that produce generic outputs or patterns that do not represent your brand if they are not explicitly counteracted.

"Do not use words like 'leverage', 'synergy', 'integrated solutions' or any generic corporate phrase. Do not start sentences with 'In the context of…' or 'It is important to highlight that…'. Do not use exclamation marks. Do not include more than three points in the list."

Those constraints seem small but they have an enormous impact on whether the output sounds like your brand rather than a generic company.

How to build a prompt library for your team

The most worthwhile investment in prompt engineering is not learning advanced techniques: it is documenting the prompts that work and making them available to the whole team. A well-designed prompt for generating email subject lines has value every time someone on the team needs to write an email. Without the library, everyone reinvents the wheel.

The structure of a basic prompt library:

  • Copy prompts: posts, emails, ads, landing pages
  • Analysis prompts: campaign data, customer feedback, competitor analysis
  • Strategy prompts: brainstorming, idea validation, editorial calendar creation
  • Automation prompts: structured JSON outputs for n8n or Make workflows

The simplest tool to get started: a Notion or Google Doc with prompts organised by use case, with the example input and expected output for each one.