Should we spend time fine-tuning a model or just focus on better prompt engineering?

All answers to this question.

For brand voice, I almost always recommend "Few-Shot Prompting" over fine-tuning first. Fine-tuning is expensive, requires a massive dataset, and makes the model "brittle"—it becomes harder to change your mind later. In my experience throughout 2023, providing 5-10 high-quality examples in the prompt (Few-Shot) gets you 90% of the way there. Fine-tuning should be reserved for teaching a model a brand-new "language" or highly technical jargon it hasn't seen before. For marketing tone, a well-structured prompt is more agile and easier to iterate on as your brand evolves.

   Answered 2025-08-25 by Heather Bennett

Have you considered a hybrid approach where you use a vector database to dynamically pull in the most relevant "voice" examples for each prompt?

   Answered 2025-08-26 by Patrick Sullivan

• This is the most modern way to handle it. By using a RAG-based approach for your marketing, you can store thousands of past campaigns and only inject the ones that match the current project into the prompt. It’s like giving the AI a memory of your best work without the cost of training a whole new model. We’ve seen this work wonders for seasonal campaigns where the "voice" needs to shift slightly from summer to winter—you just update the database, not the model.

     Commented 2025-08-27 by Russell Peterson

Fine-tuning is overkill for 95% of businesses. Focus on the prompt. If you can't describe your brand voice clearly to a human, you can't teach it to an AI.

   Answered 2025-08-28 by Raymond Jenkins

• Simple and true! Clear communication is the foundation of any successful AI strategy. We often forget that the "Prompt" is just a set of instructions.

     Commented 2025-08-29 by Pamela Reed