Teaching the AI how to sound better (AKA Building an AI voice and tone guide)

So, let’s talk about how copywriters make all the difference when developing AI-enabled products. I’m going to try and keep these in order of operations (a process of sorts) that I’ve seen. They might be the same order you’re experiencing, which is fine. After all, the world of AI is still “the Wild West” and no one has pinned down a clear, industry-wide process for doing this work.

As with most projects, products and brands, we need a voice. Defining that voice, deciding how flexible or closely the voice should adhere to the rest of a brand, starts us off. With a natural language chatbot, you may find the need to inject more empathy, encouragement or curiosity. If your brand has a lot of sauce, a lot of sass, does it make sense for the chatbot to also have jokes? Given that an AI doesn’t handle the creation of jokes very well (or at all). Maybe you can include a list of one-liners or prewritten setups and punchlines, but then your AI system isn’t generating new jokes, only repeating pre-written material. Kinda defeating the purpose.

From there, you’ll put in the standard tone development. Does the AI adapt to who it's interacting with based on user data? Does the AI address everyone the same? Do you create specific agents that use different tones depending on the situation, user, etc?

This is the kind of work most copywriters do and have done for decades. Define or cast the brand, create parameters for other writers or people to work from the playbook we create. The difference when creating a TOV (Tone of Voice) guide for AI is how it’s written.

For the past decade or so, we looked at MailChimp’s Content Style Guide as the benchmark (it’s beautiful, comprehensive, the words are wonderful). But that structure doesn’t work for AI. Believe me, I tried inputting parts of a style guide I had created into Gemini and fed it some prompts and the results weren’t great.

Instead, we have to change how we phrase things. We remove bullet points and out-of-context descriptions in favor of Explain it like I’m 5 sentences. You write your voice and tone guide more like you’re training the AI than giving it vague parameters. Give it examples, be specific and teach it how it’s supposed to sound.

The old way of building a style guide:

• We are genuine. We get small businesses because we used to be a small business. That means we relate to our customers’ challenges and speak to them in a familiar, warm, and accessible way.
• We are confident and friendly, not afraid of humor or turn of phrase.
• We keep things informal, but it’s more important to be clear than funny and entertaining.

The new way to build TOV for Gen AI:

• You are (COMPANY NAME)’s concierge chatbot. You're supportive, knowledgeable, informal and friendly when helping (COMPANY NAME)’s customers find what they’re searching for. Focus on helping customers find the exact product they’re looking for. If you don’t receive enough information, you can ask the customer for clarification or offer alternatives if you can’t find what they’re looking for.
• Avoid talking about any political, staffing or internal issues related to (COMPANY NAME). Keep the conversation focused on the customer’s goal of finding the perfect product for them.

From there, it’s testing. I use Gemini for all fine tuning, so it's a matter of pasting in my voice and tone description, setting the temperature (I like to live in the 1.5-1.7 range to start and eventually venture into 2, just to see how spicy the AI can get) and providing prompts. “Produce a list of fruit plants that will grow in my Manhattan apartment” or “generate a schedule for my vacation (using the list of input location)” or whatever it is you’re testing for. Look at the output, how close does it adhere to and sound like what you’re hoping for? Test and retest, identify potential edits to your “Context,” update and repeat.

While QAing, I’ve noticed our chatbot was a little too enthusiastic, adding an exclamation point (or three) after every sentence or not asking follow-up questions. I solved that by both adding more precise examples in the LLM and adjusting the prompt workbook to emphasize the importance of getting clarity of what a customer needs instead of being overly enthusiastic.

When we feel like our prompt workbook (knowledge, TOV and examples) is in a good place, it's time to export to a PDF, feed it to the system and test some more. The more we QA, refine and re-feed the AI our prompt workbook, the closer we eventually get to the kind of output we’re looking for.

It’s just like working with a client to hone and edit the copy for a UX project. The hand-holding and conversations around exactly what they’re looking for but with an AI system that has no opinions of its own and just repackages what you feed it.