Show and Tell
Two ways to control what AI generates.
You can describe your writing voice to AI in detail. Direct but warm. Short paragraphs. Sentence fragments when they earn it. No em dashes. Confessional. A slight Chinese-Australian inflection.
The output comes back close. Sometimes very close. But there is always a drift. A touch of over-formality. A word you would never use. An em dash sneaking back in.
Description (or tell) has its limits. Some of what makes writing uniquely yours lives below the level of any rule you could write down. You feel it when you read it. But you cannot quite name it.
So you show it.
Showing is a kind of imitation.
Paste in three paragraphs of your actual writing and say: “match this voice.” The drift tends to disappear immediately.
Same AI model. The only thing that changed is that you stopped telling the model what your voice sounds like and showed it what it actually sounds like.
This is imitation (not in a bad way), and it is the right tool when the target is too specific to describe. You want this email to land the way that one did. You want a touch of Charles Dickens. You cannot fully explain why it works, so you point at it and let the model infer.
You point at something good and say, “More like this.” The model has a concrete target to match instead of a description to interpret. No technique to memorise. You do this every time you send someone a link and say, “write it like that.”
See it in action
Say you want a LinkedIn post arguing that most corporate AI training is wasted money.
Here is the “Tell” version:
Write me a LinkedIn post about why most AI training is a waste of money. Use a direct, warm, slightly contrarian voice. Short paragraphs. No corporate jargon.
You might get an output that reads like this:
In today’s fast-paced business environment, companies are investing heavily in AI training. But are they seeing real returns?
The truth is, most AI training falls short. Employees attend workshops, learn the basics, and then return to their old habits.
Why? Because training without application doesn’t stick. Real transformation happens when people use AI to solve actual problems.
It’s time to rethink how we approach AI upskilling.
Hmm…Competent. Generic. The opening is a common filler phrase you would never write. It reads like every other AI-generated post in the feed.
Now the “Show” version. Same request, with two paragraphs of real writing pasted in first:
Here are two paragraphs I wrote recently. Match this voice, rhythm, and sentence length.
Most companies treat AI training like a software rollout. You book a vendor. Run a workshop. Tick the boxes. And three months later, nobody is using the tool any differently than before. Everything reverts to the status quo.
The problem was never the training. People did not need to learn the software. They needed a reason to care and to participate in a meaningful task that’s relevant to them. Show someone how to fix the thing that actually annoys them every Tuesday afternoon, and they will teach themselves the rest.
Now write a LinkedIn post about why most AI training is a waste of money.
And the output shifts:
Most companies treat AI training like a kind of fire drill. Everyone gets into a room. A vendor talks. Everyone files out and forgets everything by Friday.
The training wasn’t the problem. People do not struggle to learn the software. They struggle to see why it matters to them.
Give someone a real task. The report they dread. The inbox that they cannot keep up with. Let them watch how AI takes a swing at it. Eliminate the process that they hate, ease the pain. That works much better than any workshop.
Stop training people on how to use tools. Show them what the tool does for the things they already hate doing, and the learning will take care of itself.
The difference between the two is that the second one had something for AI to imitate.
One thing, though. The show version is good partly because the sample is good. Paste in a flat, generic writing, and you will get flat, generic output that matches it. The model amplifies what you give it. Give it sharp, get sharp. Give it generic, get generic.
Where each one wins
Description and showing can work together because they fix different problems.
Description sets the boundaries you can name. The rules. The things to avoid. No em dashes. Lead with the anecdote. These do not drift, because you have made them explicit. This is the control you hold on to purposely. Especially if you want to build a consistent voice and tone over time.
Showing sets the texture you cannot name. The rhythm, the choice of words, and the emotion that rules cannot be captured with a description. You point at it, and the AI model picks it up.
Description, on its own, tends to drift toward generic because the unnamed parts go missing. Showing alone leaves the model guessing which features of your sample to copy: the topic, the length, the voice, the structure. Give it one passage, and it cannot tell what you meant to point at.
Put them together, and you pin down the output. The rules provide the boundaries. The examples, the soul.
The move that makes it stick: a style sheet
If you write anything regularly, a newsletter, client reports, posts in a recognisable voice, you do not want to reconstruct this every morning. You do not want to hunt for a fresh example each time and hope it is representative.
You want the voice set once and reused. That is where a style sheet comes in.
A style sheet is a description and showing, written down together and kept on hand. The rules and guidelines are clearly spelled out. A few anchor passages that show those rules in the context of the piece. You load it at the start of a conversation with the AI, or you save it as a persistent instruction, and every output starts with your voice rather than the model’s best guess of what it should sound like.
This is the difference between negotiating your voice over and over and owning it. The reader who wants greater control over what AI produces is not looking for a cleverer prompt. They are looking for their voice to come through.
And your voice is also changing.
Here is where it gets interesting, and where most advice about “training AI on your voice” misses out.
There is no single, fixed version of your voice to capture.
Think about Dickens. There is no one Dickens style. The young writer of the early serials is not the same hand that wrote the late, darker novels. The man’s voice deepened as he did. Shakespeare’s sonnets shift across his life. Any writer you have read for long enough, you will see the change, even if you cannot say exactly when.
Your own voice is the same. The way you write this year carries things that last year’s did not. A new preoccupation. A rhythm you have grown into. A habit you have dropped.
So a style sheet is not a fixed artifact. It is a living document.
The rules tend to hold. The things you refuse to do, the em-dash, the throat-clearing opening, those can stay. But the anchor passages should refresh as your writing matures. Every few months, swap in something newer, something that sounds more like where your voice is now than where it was. The model imitates what you give it. Give it last year’s voice, and that is what comes back.
This is the reward in the whole practice. Building a style sheet forces you to notice your own voice and persona clearly enough to write them down. Maintaining one forces you to notice it changing. You end up paying closer attention to how you write than you ever did before AI. AI is like a mirror. The tool meant to imitate your voice ends up sharpening your sense of it.
Coming next
The next article tackles a habit most people never break: treating the first AI output as the answer instead of the starting point. We will look at why iteration is the actual workflow, and how to think about prompting as a conversation rather than a transaction.
Describe what you can name. Show what you cannot. Save both, and keep them current.