Maintain Brand Voice Across Thousands of Pages with AI

Maintaining a consistent brand voice across thousands of pages is a Herculean task. AI-powered solutions like GPT-4 & Claude 3.7 are changing how businesses approach this challenge. Let's dive into how these digital wordsmiths can help you keep your brand's personality shining through, even when you're churning out content faster than a caffeinated typist.

In this post we'll be introducing 3 primary techniques for getting LLMs to write in your brand's tone and style.

1. Prompting - Basic

2. Few-shot Prompting - Intermediate

3. Fine-tuning Models - Advanced

   
   

1. Prompting - Basic but necessary

Imagine having a super-smart intern who could perfectly mimic your brand's voice after just a few instructions. That's essentially what you're doing when you use prompting techniques with LLMs. By providing detailed guidelines about your brand's tone, style, and values, you're essentially giving the AI a crash course in "Speaking Your Brand."

Writing a 'Brand Content Guidelines' document

Telling the AI to "write in a confident straight-forward tone" wouldn't do much, you'll likely still see it spit out some garbage AI content. Instead you're instructions should be a lot more elaborate.

Here's what you should include in your instructions:

1. Overview of the Brand: Provide a quick summary of the brand, including its mission and vision.

2. Target Audience

• Demographics: Describe the age, location, gender, and any other relevant demographic details of the target audience. The target audience should be separated out by product or segment if the company has multiple products.

• Psychographics: Discuss interests, lifestyles, and behaviours that characterize the audience.

3. Voice and Tone

• Brand Voice: Describe the personality of the brand as if it were a person. How does it speak? What are its characteristic traits?

• Tone: Outline the emotional inflection that adjusts based on the content type or audience mood. Provide examples for different scenarios (e.g., professional, friendly, authoritative, casual).

4. Language and Style

• Language: Specify the type of language to use (simple, technical, academic). Mention any phrases or jargon typical to the brand or industry.

• Grammar and Syntax: Detail any preferences in sentence structure (short and punchy, long and descriptive), and grammatical constructions specific to the brand (e.g., active vs. passive voice).

• Point of View: First person (we/us), second person (you/your), or third person (he/she/they).

5. Use of Keywords: List key SEO strategies relevant to the content, including keyword density and placement guidelines.

6. Do’s and Don’ts: Provide clear examples of what aligns with the brand and what doesn’t.

We created this tool to help you generate all these content guidelines for your brand, you can then start using it in your prompts and content workflows. You just need to provide 3 example blog posts and your brand's homepage.

2. Few Shot Prompting

This is by far the best way to get LLMs to respond in your tone and style if you have a small blog that doesn't have enough posts for fine-tuning. For this approach, all you need is one good blog post example that encapsulates your brand's style.

Few-shot prompting is a technique used in LLMs where you provide it with examples of previous messages exchanged between you and the AI. Think of each prompt you send to the Chat Model as an input and everytime it responds as an output. In few-shot prompting, you would provide the LLM with input and output examples.

In Moonlit Platform, you can apply this technique in your workflows by setting the Message History field in the Chat Model function in the following format:

[{"role": "user", "content": "Write an article about x"},
{"role": "assistant", "content": "{{example article}}"}]

Then in the Prompt field you can say something like Now repeat this task for topic Y.

Now of course this is a very simple example, to improve it you'll need to include more context and information in your prompt. You can also provide it with the brand content guidelines instructions from the previous technique we discussed above as well a research about the topic with external resources for the AI to cite.

3. Fine-tuning - Training AI models on your Blog Posts

This is the most advanced and complex approach. It is only suitable for blogs that already have at least 40 post that are well crafted and optimized. This process involves creating a training dataset consisting of prompt/response pairs. The responses in this instance are your actual blog posts and the prompts should be the instructions you would give to a writer to output such as blog post.

Now of course, it's an extremely tedious process to create this training dataset manually so we recommend you check out our Personas feature which allows you to create these fine-tuned models in minutes.

We've tried this ourselves for multiple client blogs and the results have been incredible. Within the Personas interface you can also quantitively assess the model's performance as it give you some test posts from your blog with a prompt to test the semantic and stylistic similarity of the fine-tuned model to the actual post and the base (untrained model) to the actual post.

Other Techniques to experiment with

The 3 techniques above are what plays the biggest role in getting LLMs to response in your style and tone, but worth noting 3 more that are easy to explore but the results would vary depending on the task context at hand.

Temperature Check

When it comes to LLMs, the temperature parameter is about how spicy you want your content. The temperature parameter in LLMs controls the level of creativity in the generated content.

• Low temperature: More conservative, predictable content. Great for maintaining a consistent voice but might lack pizzazz.

• High temperature: More creative, varied content. Excellent for brainstorming but might occasionally go off-brand.

Finding the right temperature for your brand is crucial. With Moonlit Platform, you can experiment with different settings to find that Goldilocks zone where your content is consistent yet engaging.

Experiment with different models

The LLM race is heated and there's a lot of different models currently on the market that have great performance. Some models might work better for your writing task. At the time of writing this article the best models are currently:

• Anthropic's Claude 3.7

• OpenAI's o3 Mini

• Gemini 2.0

• DeepSeek

• Qwen 2.5

All of these models and more are available on Moonlit Platform so you can easily build your content workflows and switch between these models to see which yields the best results.

Banning Words

This is a very common approach taken by SEO and content specialists to avoid words commonly used by AI. If you've had decent experience using AI for content, you're probably familiar with the cheesy intro lines it can commonly throw like:

In today's digital age…

Or

In the rapidly evolving landscape of…

So in your prompts you can instruct the AI to avoid using certain words or phrases to keep it more human.

Final Thoughts

Maintaining a consistent brand voice across thousands of pages is no longer a pipe dream. With AI-powered tools like Moonlit Platform, you can ensure your brand's personality shines through in every piece of content, whether you're creating a single blog post or an entire library.

With the proper prompting techniques, finding the right temperature settings, creating comprehensive guidelines, and using the Bulk Runs feature, you can create a content strategy that's as efficient as it is effective. So why not give your brand voice the AI amplification it deserves? Your audience (and your content team) will thank you.

We went through a wide array of methods in this post and some might be more confusing than others, if you would like to learn more about implementing any of these techniques in Moonlit Platform, feel free to schedule a call with our team.

‍