My marketing agency is struggling to keep a consistent "witty yet professional" tone across 50 different client accounts. Can we solve this through prompt engineering by creating a standard Persona Prompt, or do we need to fine-tune a custom model for every single client we manage?
3 answers
Fine-tuning is usually overkill for brand voice. Instead, invest your time in high-quality prompt engineering for "Persona Definition." We create a 200-word description of the brand’s "personality," including what words to use and what topics to avoid. Incorporate a "negative prompt" section like "Avoid using corporate buzzwords like synergy or leverage." In early 2023, we moved to this template-based approach, and it allowed us to scale our content production significantly while keeping each client’s unique voice intact across all social media platforms.
Do you provide the AI with examples of past successful posts? Prompt engineering is much more effective when you include 3-5 "gold standard" examples of the voice you want.
Lowering the temperature is a great idea. It makes the prompt engineering work harder by forcing the model to pick the most likely "on-brand" words.
Michael is spot on. I’ve noticed that temperature 0.3 paired with a strong persona-based prompt engineering strategy is the "sweet spot" for most of our B2B clients.
We do use examples, Bradley, but sometimes the AI gets "too" creative and strays from the script. I'm wondering if we should adjust the "temperature" setting in the API alongside our prompt engineering to make the output more predictable? Our current temperature is 0.7, but maybe lowering it to 0.4 would help keep the brand voice more stable and less prone to random jokes.