How to Make AI Write a Proper Article

I asked an AI to write something for this site about AI tutors. What came back looked finished. Headings. A neat thesis. A table. Ten rules. All the furniture of an article and none of the life.

It was the sort of thing you skim and forget before you reach the end. No family in it. No kitchen table. No sense that a real person had ever sat with a stuck child and felt the particular mix of patience and panic that is home education. Just clean paragraphs that could have been bolted onto any edtech blog in the world.

That is the default. Not because the machine is evil. Because the machine is a people-pleasing summariser of the internet’s average voice, and the internet’s average voice on “how to use AI in education” is managerial sludge. If you do not stop it, that is what you get. Then you publish it, feel vaguely ashamed, and wonder why nobody cares.

This piece is about not doing that. Not a list of magic spell prompts. A way of working that treats the AI as a junior research assistant with no taste and no memory of your life — which is exactly what it is.

Why the first draft almost always dies on the page

Read a bad AI article out loud. You will hear the tells before you can name them.

Everything is balanced into threes. Every section has a bold claim, then three supporting bullets, then a pivot sentence. Nobody talks like that at the dinner table. Real people digress. They remember a name. They get annoyed. They admit they are not sure.

There is no place. No weather. No child who hated fractions on a wet Tuesday. Abstract parents and abstract children float in abstract homes.

There is no cost. Real advice has a price: time, pride, the row you have when you take the laptop away. Dead prose never risks a fight.

And there is always the posture of the helpful pamphlet: “Here is the hierarchy.” “Here is the spine of the piece.” As if families run on laminated posters.

If your draft could be pasted onto a startup blog without changing a word, it is not ready for a site that has been talking to real home educators since the dial-up years. Ours has. We know the difference between a resource and a brochure.

What you actually want from the machine

Be honest about the job.

You do not want the AI to be the author of record. That way lies hollow pages and a slow rot in your own ability to think on paper.

You do want help with the boring scaffolding: structure options, counter-arguments you forgot, a brutal list of what a sceptic would say, a plain-English explanation of a technical bit, a reminder of related ideas already on your site. Sometimes a first splash of words you will mostly throw away.

The moment you treat “write me the article” as the whole brief, you have already lost. That prompt invites average. Average is what the training data is full of.

So change the job description in your head: I am writing this. The machine is interrupting me helpfully when I ask.

That sounds slower. It is. Good writing is slower than generating sludge. Homeschooling taught most of us that already.

Before you type a single prompt

Do this on paper or in a plain note. Five minutes. Not optional.

1. One true story.
Something that happened. A name if you are willing. A room. A failure. If you cannot find one story, you do not have an article yet — you have a topic. Topics are what content farms sell. Stories are what people remember.

2. One sentence you actually believe.
Not a marketing thesis. The thing you would say to a friend over tea if they asked. Rough is fine. “AI tutors are fine for drill and deadly if the child never struggles.” That kind of thing.

3. Who it is for.
A tired mum in Reading. A dad who is good at woodwork and scared of algebra. A Christian family who will not hand formation to a chatbot. Name them in the brief. If you say “homeschoolers” and stop, the machine will write for a demographic spreadsheet.

4. What you refuse to sound like.
Corporate. Guru. Fearmonger. Cheerleader for whatever Silicon Valley is selling this quarter. Put the bans in writing. The model will still drift toward them; you will catch it faster if you named the enemy.

5. Where this lives.
On this site: Christian home education, parents responsible, real families, tools not masters. Paste a short scrap of your own old writing into the chat when you start — a paragraph with your rhythm — and say: match this human, not the average blog.

Without those five, you are asking a stranger to guess your soul. They will guess LinkedIn.

How to talk to it so it cannot phone it in

Do not say this

Write a comprehensive article about AI tutors for homeschoolers. Include benefits, risks, and best practices. SEO friendly. About 2000 words.

You just ordered the sludge special. “Comprehensive” means padded. “Best practices” means committee language. “SEO friendly” means written for a robot to rank, not a mother to trust.

Say something closer to this

I run a Christian home education site. Voice is plain, first person when it fits, not corporate, not hype. I already have older pieces that sound like a real parent talking — imperfect grammar is fine; laminated “10 rules” posts are not.
Topic: AI tutors. My real position: useful for drill and re-explaining after the child has tried; dangerous when it steals the struggle or pretends to be a moral guide. I care more about formation than features.
Here is one true scene from our life: [your scene — even a short one].
Do not write the finished article yet. First: ask me ten sharp questions you need answered before anyone should write this. Questions about our family, what we already tried, what I will not compromise on. No generic education questions I could pull off a government PDF.

That last part matters. If it never asks you anything, it is guessing. Guessing is how you get “parents should establish clear boundaries” for the thousandth time.

Answer its questions in your own words. Messy answers. Half-remembered. “We tried it with Julian on physics and he stopped opening the textbook.” That mud is gold. Paste it back in.

Then:

Now give me three different openings — a paragraph each — in my voice. Not outlines. Not titles. Openings. One should start from the scene I gave you. One can be angry. One can be quiet. No bullet points. No “In today’s rapidly changing world.”

Pick one opening that does not make you wince. Or steal a sentence from each and write your own. Then continue section by section, not as a 2000-word dump. After each section you say what is fake, what is missing, what only you know.

When it gets preachy or triangular, say so in plain language:

This sounds like a DeepSeek brochure. Kill the numbered hierarchy. Put the child back in the kitchen. Shorter sentences mixed with longer ones. One awkward honest line I would actually say out loud.

Cruel feedback works better than polite feedback. The model is not your friend. It is a tool that over-complies with softness and under-delivers on taste.

The checks that catch lazy output

Before anything goes near your site:

Read it aloud. If you stumble from boredom, readers will leave. If every sentence has the same weight, rewrite until some are short enough to sting.

Hunt the tells. Delete or rewrite: “In conclusion,” “It is important to note,” “In today’s world,” “leverage,” “landscape,” “delve,” “robust,” “empower,” “journey,” “whether X or Y, the key is Z.” And those neat little triad stacks. Real writers use lists sometimes. They do not build the whole house out of them.

Find the blood. Is there one sentence only you could have written? A name, a place, a regret, a joke, a belief that would get you uninvited from a polite education conference? If not, it is still generic. Add the blood yourself. Do not ask the machine for “more authenticity.” It will fake that too.

Make it doable. Abstract virtue (“keep the parent in charge”) is useless until you say what happens on a Tuesday. Who opens the laptop. Who sets the task. What the child must finish first. What you do when they paste the whole worksheet into the chat. If you cannot picture a family doing the advice by Friday, rewrite until you can.

Check truth like you mean it. Models invent sources, soften hard claims, and borrow other people’s certainty. Anything factual — get a second source. Anything theological or moral — that is your job, not the model’s training data.

Sleep on it. The glow of “look how fast we made 1800 words” is a trap. Morning you is a better editor than midnight you.

A small workable rhythm (actually doable)

Not a lifestyle system. A rhythm for one article.

Day 1, twenty minutes. Story, belief sentence, audience, bans. No AI yet.
Same day, fifteen minutes. Prompt for questions only. Answer them in rough notes.
Day 2. Openings. Choose or rewrite one by hand.
Day 2–3. Draft in chunks. You write the hard paragraphs. AI helps with the mechanical ones if needed.
Day 3. Read aloud. Cut a third. Add one true detail you left out because it felt small. Small is often the point.
Publish when it sounds like someone in your house could have said it.

If that feels too slow compared with “generate full article,” good. You have noticed the trade. Homeschooling is full of those trades. We already chose the slow road for our children. Our words can take the same road.

Why bother — the heart of it

You might ask why a home education site should care about prompting technique at all. Because our children are watching how we handle powerful, lazy-making tools. If we paste machine prose onto the web and call it wisdom, we teach them that appearance beats substance. That is the same lie government schooling sells with worksheets and grades.

Writing is thinking made visible. If a machine does the writing, someone still has to do the thinking — or nobody did. Parents who take responsibility for a child’s mind should not outsource their own voice the first time a chatbot offers a tidy outline.

There is also love in this, oddly. A proper article on a small Christian site is a letter to someone you may never meet: the parent up too late, worried they are failing, searching for a word that steadies them. They can smell when nobody human was home in the prose. They deserve better than a well-formatted shrug.

So use the machine. I do. For sorting, for sparring, for “what am I not seeing?” For a kick when the blank page wins. Then take the words back. Put your family’s fingerprints on them. Leave a few rough edges. Tell the truth you would stand by if the internet went dark tomorrow.

That is how you make AI write a proper article.

You don’t.

You do — and you refuse to let the tool publish in your name until the thing on the screen could only have come from your life.