Deepseek is red-pilling me on creative copywriting

Jan 27, 2025 · Ravdeep Singh

Creative copywriting looks at existing copy on a landing page and suggests a better, personalized version of it. It's an interesting problem because there's no one right answer. Nike, Supreme and Zalando all sell shoes but take a different approach to presenting themselves. The same way JK Rowling and Joan Didion are both authors but hard to imagine them writing a similar book.

There's a minimum expectation of grammatical correctness from an output but beyond that, it's two things which are hard to automate:

  1. Personality
  2. Taste

Personality is having your own voice built over a few decades and Taste involves editing to bring across unique insights. Extending the shoes example. Nike's brand personality is to celebrate athletes and taste involves getting the point across by celebrating paralympians.

In the last few months, I've upgraded the LLM model powering this problem a few times. This journey started with llama (3.1 then 3.3), then x.ai's grok and most recently gemini. The output was always decent but never good enough.

There's no mathematical way (that I can find) which rates quality so it's always a human copywriter rating responses. I've tried automating this with one LLM writing and another rating the response but then it's subjected to evaluator LLM's biases for what is good?

With deepseek (or broadly, reasoning models), the thought process breakdown is simply magical. Enough for me to be confident that personality is nearly solved while taste might still be some distance away. For languages I speak natively, I cannot beat it anymore. Few examples:

ControlBrandDeepseek
United by rhythm. Discover the new Black Unity watch band and face.appleCelebrate unity through rhythm with the new Black Unity watch band and face.
Find something special for your special someone.appleDiscover the perfect gift for your special someone.
wöchentlich an Deine Haustür gelieferthellofreshJede Woche frisch an Deine Tür
Essentials for Every MilenikeMust-Have for Every Journey

What hits you in the face is the reasoning it goes through to craft something unique with a consistent tone. In real-time it walks through the thought process with a simple prompt and answers "Okay, so I need to rewrite the user's input to make it suitable for an AB test as a CRO specialist...". The process improves the RAG and the initial prompt so much more that I've rebuilt tooling around reasoning models to make this approach more future proof.

Long-winded way to announce a product feature but adding a footnote. Wafrow chrome extension and creative copywriting within the dashboard are now running on deepseek. Please give it a spin and let me know what you think.

I usually avoid crystal balling but this feels like a pretty seminal moment. Arguably, one marketer with access to such an LLM will run circles around a team of marketers operating in multiple countries.


p.s. On principle, I still refuse to read a book written by an LLM. If you didn't bother writing it, why should I bother reading it? I wonder how long this resistance will last. What if an LLM authors better sci-fi? What if?