In the British Museum you’ll find a six-sided prism made of baked clay, about the size of a large wine bottle, and covered top to bottom in tiny script.
It’s a nearly 3000-year-old artifact called the Taylor Prism, and it starts like this (translated from cuneiform):
“Sennacherib, the great king, the mighty king, king of the universe, king of Assyria, king of the four quarters, the wise shepherd, favourite of the great gods, guardian of justice, lover of righteousness…”
It goes on. At length. And there’s only really one word to describe it: slop.
Slop isn’t the preserve of AI. Humans have been generating it for millennia. And the root cause is the same: lack of thought – either about the audience or the task.
The fix isn’t a better model or a newer tool. It’s more friction.
Let’s get frictionmaxxing
In marketing, we spend our lives sanding friction away. But deliberately introducing points of care and intention into the way you work with AI – frictionmaxxing – is the difference between getting real value and producing an endless stream of meh.
Here are five ways to do it (but there are thousands):
1. Evaluate models against things that actually matter to you
Have you heard of FrontierMath Tier 4? OSWorld-Verified? CyberGym?
How about lighting and shadow, photorealism, composition?
The second group are the metrics our design team uses to score image models. And that’s the point. When a new model drops and the internet loses its mind over its ARC-AGI-3 ranking, that tells a design studio precisely nothing about whether it’ll produce imagery they can work with.
So do your own evaluations. Run the same brief through a number of models, then score the outputs against things your business cares about. Keep a record, and re-run the evaluation every time a new model is released.
2. Prompt architecture: different media, different direction
Image prompts, video prompts and audio prompts aren’t the same thing. Most people treat them as if they are, then wonder why the output feels generic.
For a visual, specify typography, define the canvas hierarchy, describe the visual language. For a video, you’re writing a director’s brief: cinematography, subject, action, context, style and ambiance. For a voiceover, think about audio profile, scene-setting, director’s notes – and don’t underestimate the power of audio tags. There’s a world of difference between [whispers] and [sarcastically, one painfully slow word at a time].
3. Know your message before you ask AI to say it
If you need to generate some writing, first think through these three questions:
- What do I want my audience to know?
- How do I want them to feel?
- What do I want them to do?
The Know/Feel/Do framework isn’t new or clever. But it forces you to own your message before you outsource the words. If you don’t know what you’re trying to say, the AI doesn’t stand a chance.
4. Prompt chaining: one thing at a time
Say you need to analyse some data, pull out five key takeaways, draft a customer email and write a subject line. The temptation is to ask for all of it in one go.
Don’t. AI takes shortcuts when you overburden it. If it cuts corners on step one, every step after that is built on a shaky foundation. Break the task into a chain of individual prompts instead. Let the model do one thing well, then hand it the next. It means you can fix any errors along the way, and there’s loads of evidence that the result will be better.
5. Modelmaxxing: the right model for the job
There’s no single best AI model. The model you use for long-form copy probably isn’t the model you should use to make videos. And a model that’s good off the shelf might be even better fine-tuned on your brand’s specific visual identity.
So if you want the best AI results, make sure you use the right model for the task at hand. This is the problem with being locked into a single subscription for Copilot or OpenAI, for instance. And it’s why – shameless plug alert – we built our own AI suite with all the best models in the same place.
The bigger point
None of this is really about AI. Sennacherib’s scribes didn’t think about what their audience needed to hear. The question that cuts through both – whatever the tool, whatever the century – is a simple one: do you actually care what comes out?
If you do, it’ll show in the output.
Written by Nick Padmore, Head of Language.
