How to make AI stop writing horsecrap

Why your prompt needs less “please write this nicely” and more “fight me in the car park.”

So OpenAI is rich as.

Anthropic is rich as.

Every day they are harping on about some other great thing that they have done.

And yet….

What the hell is going on with their writing department?

Do they have a budget of like $50 or something?

Here is an example.

It’s a pretty basic request. Just tell me why you suck at writing.

But can the trillion dollar super magician AI company do that?

Absolutely not!

This is the drivel I get back.

ChatGPT sucks at writing because it is trained to sound like writing before it understands why writing matters.

That is the basic problem.

Good writing is rarely just competent. Good writing has blood in it. It has a person behind it. It has pressure. It has judgement. It has taste.

The result is competent. That is the insult.

Huh?

I swear ChatGPT 2 could write better than this horsecrap.

I swear ChatGPT negative 487 could have written better than this.

Do you know the annoying thing about all of this horsecrap?

We are forced to read this horsecrap every single day.

Whether it is cover letters.

Or request for proposals.

Heck, even my junk email has people sending the exact same thing.

And while I have high self esteem, I know that Cub Digital has done absolutely nothing to afford Naya and Dolores reaching out from different companies at precisely the same minute to purchase us.

Get out the red pen.

I’m an optimistic person. I’m a happy person. I’m not an angry or bitter person. But I do have hayfever and boy is that the lamest allergy ever invented. And just like when you have hayfever and you go to an Airbnb and you start sneezing and everyone asks are you crying and you say no I’m fine, it’s just the dust, that’s what happening with me with AI writing. I’m crying and sneezing and my eyes are red.

This is all a long way of saying that I miss Ernest Hemingway.

He would have somehow figured a way to punch OpenAI & Anthropic in the face at the same time for this.

But, besides my allergies and aversion to all of this, there is something deeper going on.

Just like in Titanic when the lookout says “Iceberg dead ahead” - well that’s me and you are the person that is going to smash into the iceberg.

Because if we can’t even automate writing or an email or an essay, then how on earth are we going to automate your business? If you are wanting this tool to somehow solve your customer support or answer the phone or send automated emails, then don’t we need to make sure that it doesn’t sound like an utter space cadet? Let alone getting it to help you with your strategy or business plan.

So in this essay, I’m putting my Erin Brockovich hat on and getting to the bottom of this.

I’m going to teach you how to stop producing utter horse crap.

First - here is why it produces utter horse crap

ChatGPT does not see your question like a person sees a business problem. It sees something closer to a sequence of tokens that activate patterns from training.

Basically when you type words like this.

• “business strategy”

• “founder”

• “advice”

• “growth”

• “risks”

• “frameworks”

• “positioning”

• “AI”

• “consultancy”

It predicts the next useful token, over and over. Inside that process, the model forms rich internal representations, but it is not naturally sitting there thinking:

“What is the weirdest but most useful angle here?”
“What domain am I missing?”
“What would a historian, game theorist, designer, military strategist, VC, therapist, and operator each see?”

Unless you force that behaviour, the model tends to collapse towards the centre: competent, safe, familiar, semi-obvious (aka utter horsecrap).

The problem is not only that the AI is trained to give you something you “like”. It is that your question often contains an invisible default: “give me a coherent answer”. So the AI gives you a coherent answer and coherent answers are often intellectually boring.

What we need to do is to request to expand the search space before answering.

For example, instead of:

What should my business strategy be?

Ask:

Before answering, explore this problem through 12 lenses: game theory, pricing, incentives, distribution, talent, bottlenecks, category design, operational leverage, customer psychology, capital allocation, technology shifts, and failure modes. For each lens, give me one non-obvious implication. Then synthesise the 3 strongest strategic options and argue against each one.

A useful mental model I use is this.

Weak prompt:

Question straight to plausible answer

Better prompt:

• Question

• possible frames

• contradictions

• weird analogies

• options

• trade-offs

• answer

And if you still need more

• Question

• explore the terrain

• identify hidden assumptions

• generate opposing models

• test against reality

• choose a strategy

Here is an example query, pretending I am running a civil engineering business.

I run a New Zealand civil engineering consultancy. 

Do not give me generic strategy advice. 

First, identify the obvious answer and explain why it is probably mediocre.

Then analyse the business through these lenses:
1. agency economics
2. productised services
3. AI substitution risk
4. founder bottleneck
5. sales narrative
6. hiring leverage
7. enterprise trust
8. pricing power
9. repeatability
10. defensibility

For each lens, give me:
- the common advice
- the sharper advice
- the uncomfortable truth
- one non obvious implication 

Then give me a blunt recommendation.

This gives a really rich response straight away.

For the civil engineering company it says that clients don’t really want civil engineering but rather they want permission, confidence, risk reduction etc.

This is actually a profound insight and then it leads to a non obvious implication that the civil engineering sector could learn more from.

Council Objection Pre-Mortem is the outcome a client wants and yet engineers are busy selling land development or stormwater mitigation reports.

One more example is in the sales section it basically critiques all civil engineering messaging.

And gives you tactical and non obvious implications of what to do with that.

That was just 2 sections, but it did that for the 12 that I asked for. By forcing the AI to explore different spaces before answering, I got a much better response.

When you want more intellectual output, use words like:

“non-obvious”
“from first principles”
“argue against yourself”
“give me the uncomfortable version”
“what would I be missing?”
“what is the boring answer?”
“what is the dangerous but potentially correct answer?”
“compare five mental models”
“show me the hidden structure”
“do not optimise for politeness”
“make the answer more like a strategy memo than advice”

Keep a list of these words and constantly add them to every prompt where you want to go deeper.

Make the model less agreeable

An interesting finding I had in my research is that intellectual work requires productive hostility.

The reason AI writes such horsecrap is because the models are agreeable. But that agreeable feature is why we enjoy using AI for therapy. It makes us feel heard.

So to counter that, we need to specifically request it.

I want you to challenge my premise. Assume I am intelligent but probably missing something. Do not reassure me. Find the weak points in my thinking.

Often if we are asking for a plan, the AI might need to tell us that our business sucks and that we are going to go bankrupt. That’s what a good business mentor would do. They know when to tell you what you are doing is wrong and not to always tell you that you are doing a good job.

Change modes to imagery for different thinking

I’ve been finding myself doing this a lot more.

If you ask the AI to generate your strategy pictured as a top down WW2 map, you get quite a different expression of the problem and I think a more interesting thing to digest that a written response.

I like this because it forces the AI to condense everything into one graphic and images are processed differently to words on a page.

It has to think about the sizing and the weight and what it makes important.

Why “do not answer yet” works

Transformers generate token by token.

Once the model starts producing an answer, it tends to continue in the direction already established. So whatever the early tokens are can set the frame and then the model can basically go AWOL on you.

If the first paragraph sounds like generic business advice, the rest often follows that path.

When you say:

“Do not answer yet. First generate possible frames.”

You are delaying the commitment of the model.

That matters because then the generated text becomes part of the context. The model can condition later tokens on earlier generated tokens.

Most horsecrap comes from the fact that you made the model answer you too quickly. When you see horsecrap out there, think to yourself, I need to slow down and let it explore.

The Ironic Conclusion

I’d love to tell you that at the end of this that if you do those prompts I did above and then get it to write an essay that it will turn out to be great, but that doesn’t happen.

Unfortunately it still produces horsecrap as soon as you get it to put it into an essay or article form.

I think the word essay has just been butchered to death by the trillions of uni essays and now you can’t ask for an article, or an essay or any of those words without it divulging to utter rubbish.

Maybe ask it for a proclamation or a declaration or a rap. Again all of those will produce a more interesting response as they are outside of the norm.

But the hope is that you got given enough ideas from the AI that it allows you to then piece together something good and write something that at least will be better than what you started with.

And with that - here is a poem to help you remember how to stop writing horsecrap

Before you ask the magic box
to “write a thing with flair,”
grab it by its token-neck
and drag it from the chair.
Say: “No warm bath of clever sludge,
no LinkedIn monk in socks,
show me frames and fights and flaws
before you touch the box.”

Do not let it trot straight out
with polished stable crap.
Make it cough up weird ideas,
then slap them on the map.
First the mess, then judgement comes,
then blood, then taste, then bite.
If AI starts sounding pleased with itself,
stop the horse … rewrite.

Bookclub

I quite enjoy these books that are psychology meets branding meets basic things. If you are a fan of Rory Sutherland then you will enjoy this book. It talks about the 16 1/2 biases that influence what we buy. Even that 16 1/2 is a bias, it sounds more precise, which makes you believe that the author is more credible (the power of precision).

View the book on Amazon here.

I’m an Amazon affiliate, if you make a purchase, I get a tiny commission.

Other Articles this week

What AI Does to Novice Coders & Why I Don’t Agree That It’s Bad