There’s no denying that artificial intelligence (AI) is driving rapid change in numerous areas of business, including digital marketing. As part of this transition, content writers are having to become AI prompt engineers, turning to large-language models (LLMs) such as Open AI’s Chat GPT and Google’s Gemini to boost their output and effectiveness.
To fully realise the potential of an LLM, it is critical to learn how to create effective prompts. A nuanced approach, especially when striving for a thorough and high-value response, yields more effective results.
Establishing the ‘writer’ persona in your AI prompt
Every experienced copywriter knows that understanding the voice of the content is key. For your AI prompt, this involves crafting an initial paragraph that communicates the expected level of complexity and nuance. It should also identify the target audience.
Some key points to consider:
- Level of complexity and nuance: Are you aiming for a beginner-friendly explanation or a more in-depth analysis for industry professionals?
- Audience demographics: Who are you targeting? Understanding your audience’s background and knowledge level helps the LLM tailor the language and references accordingly.
- Desired tone and register: formal, informal, humorous, or authoritative? Specifying the desired tone ensures the LLM adopts the appropriate style.
For instance, when writing a B2C blog, you might specify that the AI adopt the role of a “seasoned commercial content writer” targeting a UK audience with content pertaining to the client’s specific range of products and services.
An example:
“We are going to work on creating content in British English for a business-to-consumer blog that is targeted at the UK market. You will act like an accomplished and skilled commercial British content writer in order to do this.”
This contextual information not only sets the tone but also subtly instructs the AI to adopt a style that is familiar and appealing to the target audience and their expectations.
The power of context
LLMs excel at producing content when given rich context about the client, their industry, target audience, and specific objectives of the content. This depth of detail about the expected outcome allows the AI to tailor its language, focus on theme, and potentially even offer insights that align more closely with the client’s brand.
The initial time investment may seem counterintuitive, given that the objective of leveraging AI is to save time. However, it will pay in the long run as it provides a foundation for streamlining future content, thereby saving time and enhancing efficiency.
The more context you provide in an AI prompt, the better the model can understand the specific task at hand. This includes:
- Client information: Briefly define your client, their industry, and the services they provide.
- Topic and desired outcome: Clearly define the topic and the specific outcome you’re aiming for. Are you looking for an informative blog post, a persuasive landing page, or a catchy social media caption?
Defining AI prompt parameters
It’s essential to remember that LLMs are ultimately tools. While they automate certain aspects, their potential is maximised when given clear ‘mental frameworks’ to work with.
Clear and precise parameters reduce the risk of vague or generalised outcomes and guide the AI in generating content that both resonates with the intended audience and fulfils the client’s objectives.
Insufficient or vague instructions increase the likelihood of the LLM making assumptions to fill in the gaps, potentially leading to incorrect or irrelevant content. When a model does this, it is said to be “hallucinating”.
The parameters you set in the AI prompt, such as title, format, structure, and tone, guide it towards producing content that meets a specific set of criteria. As you iterate and tweak your prompt for future content, each parameter can be adjusted to align with different content goals.
LLMs understand natural language commands and have been trained on an enormous amount of publicly available data, so even if you’re not sure exactly what you want at the outset, giving the model comparative statements can still help set it in the right direction, e.g., “I want the document to read like a professional white paper” or “I want the post to have the feel of an article you might find on Substack.”
Some additional aspects to consider:
- Length and word count: Be specific. Give the model a number or number range to aim for in terms of characters, words, sentences, paragraphs, or pages.
- Comprehension level: Tailor the complexity of the language to your target audience’s understanding, e.g., “Keep the reading comprehension level to the ability of a typical secondary school leaver; it should not exceed undergraduate level.”
- Regional variations: If your content is intended for a specific region, specify the dialect or that which you don’t want replicated, e.g., “please avoid using US-centric references.”
Adding context and parameters to your AI prompt
Specific, actionable guidance enables the AI to perform optimally. The early sections of a prompt for article generation should provide plenty of nuance for the model to work with and interpret:
“The goal is to offer detailed coverage of the topic by providing information that fulfils the search intent of the reader, offering a detailed standalone guide suited to the core audience.
“The article you write should assist the reader in understanding the topic by providing accurate and clear copy that makes effective use of on-page SEO strategies to inform, explain, and describe in an authoritative style. The article must be able to convince the reader that the author has secured practical experience with the topic.”
This may seem excessive, but it means the model can more precisely meet a certain set of demands.
For example, because the AI has the ability to read and comprehend information in its training data relating to effective SEO, it can apply these principles in its writing when explicitly (or implicitly) instructed to do so. Keyword terms such as “search intent” and “on-page SEO” provide context that can help the model understand that it should try to engage in an SEO-compliant approach.
Obviously, this is dependent on the quality and depth of the model’s training data, so it’s not correct to assume that simply asking an LLM to do something based on a body of knowledge would always result in the desired output. You must experiment and evaluate the outcomes yourself.
Structuring for depth
Specify the desired structure at word-, sentence-, and paragraph-level to fine-tune the shape of the output:
“I’d like you to utilise structured headers (mainly H2 headers) to divide your ideas into sections. However, these sub-topical portions should always build on an initial idea over several paragraphs.”
Conciseness is beneficial, but not at the expense of depth. Explicitly request detailed explanations to ensure that the model does not prioritise brevity over providing useful, comprehensive information.
Encourage the LLM to expand on ideas rather than seeking the most concise solutions. If output is short, ask it to refine its output, e.g. “add additional steps to your logic when considering X,” “provide more examples about Y,” “elaborate on point Z, taking into account A, B, and C.”
Controlling voice and complexity
Directly instruct the LLM on your desired tone and language and explain how this could impact understanding:
“Your draft should have a formal tone but not be too academic. It must be appropriate for a general readership and not be intended to be exclusive due to formality, esoteric references, or intricacy. Try to keep jargon to a minimum, but don’t be hesitant to use technical terms when appropriate.”
Learning from examples
Providing the AI with samples of the desired content style can have a significant impact on the tone, structure, and level of information in its output.
This could entail incorporating an excerpt from a prior piece that effectively represents the client’s brand voice and content requirements, even if the topic is different.
For example:
“To assist with getting the tone, length, language, and structure right, I’m going to supply you with an example of content written for the client, but on a different topic.”
This approach not only gives the AI a concrete reference for the desired style but also encourages consistency across different pieces of content, enhancing the overall coherence and effectiveness of the client’s content strategy.
The importance of the human touch
It’s crucial to note that even the most advanced LLMs can’t replace the human touch entirely. A good copywriter brings strategic understanding, nuance, and the ability to connect emotionally with the target audience. LLMs work best in tandem with human expertise.
- Editing and refinement: Treat AI-generated content as a robust first draft. Fact-checking, refining structure, and injecting your unique brand voice will still be essential.
- Strategic thinking: LLMs excel at language modelling, but the overall content strategy, ideation, and the ability to see the ‘bigger picture’ still lie squarely in the hands of the human content writer.
Key considerations in building your AI prompt
- LLMs require careful instruction to maximise their value. View the model as a tool that thrives on specific, detailed input and guidance.
- Initial time investment in prompt refinement pays dividends in long-term efficiency and quality.
- LLMs are trained on large datasets, but their knowledge can lag behind real-time information. This can result in suboptimal outcomes if current data is required for a specific activity, such as keyword research for an article.
- The technology behind LLMs is constantly evolving. Experimentation is key to discovering the most effective ways to integrate these tools into your workflow. Be prepared to learn, adapt, and refine your prompts over time.
Taking things further
This exploration has highlighted the power of nuanced prompting when working with LLMs. In the next instalment, we’ll examine:
- further prompt refinement techniques
- strategies to ensure content quality
- the advantages of utilising multiple LLMs
