The Human Side of Technology
The Public Guide for Writing with AI Collaborators
There is a very old piece of advice for writers:
Write what you know.
I think it needs a small update.
Write what you know or what you'd like to know.
That's where nearly everything I write starts. Not with an AI prompt. Not with a blank document. Usually not even with the intention of writing something.
It starts with an observation.
Something catches my attention and I ask a question about it. Sometimes I make a statement about it, if only to myself. Most of those ideas are sparks. They show up, make a little light, and disappear.
Others become ear-worms.
They roll around in my head while I'm doing something else. I pull at them. I ask more questions. Sometimes I get distracted and lose part of the thought. If it comes back, though, I usually get more out of it than whatever I lost.
By the time I bring one of those ideas to an AI collaborator, I've often already spent hours with it.
That's important.
Because the prompt is almost always not where the thinking started.
1. The idea is yours before AI ever sees it
Sometimes I write because I think I know something worth sharing.
Sometimes I write because I don't know something and want to work it out.
Sometimes I want to pull at the threads of an idea and see whether it survives.
Sometimes I write because I want to become a better writer, and there is only one known cure for that.
And sometimes I write because I want to leave something behind.
The common thread is intention.
If you're going to write with an AI collaborator, have an idea first. It doesn't need to be a fully formed thesis. A question is enough. An observation is enough. Curiosity is enough.
But bring something to the table.
Otherwise, you're not collaborating on your idea. You're asking something else to give you one.
Those are different activities.
2. Follow the idea somewhere
At my core, I'm a scientist and an engineer.
We make hypotheses.
We should not have expectations of the results.
That's harder than it sounds, particularly when you've already fallen in love with your clever idea.
I once followed a seemingly simple statement about monkey saddles in geometrical space far enough that it challenged me to become better at mathematics and to understand that what I was seeing might never have been seen quite that way before.
That's the fun part.
Follow the path.
Research. Ask questions. Look for the thing that proves you wrong. If credible evidence challenges your thesis, don't hide it. The challenge may be as important as the thesis.
I usually know I'm ready to write when a fundamental understanding of the evidence becomes nearly clear.
Not certain.
Clear.
We would all be a bit better off if we freely admitted that we might not be right about something.
AI doesn't change that rule.
If anything, it makes the rule more important.
3. Talk it out
Feynman gets dragged into a lot of conversations about learning, but there's a reason for that.
If you can't teach it, you probably don't understand it well enough yet.
I've always encouraged discussion among my engineering collaborators. Put smart people around a problem and make them explain what they think. Challenge each other. Ask the question somebody else hasn't asked yet.
It is how we all get better.
Writing works the same way for me.
I like to talk it out.
Sometimes that's with another human. Sometimes it's with several humans. Sometimes it's with AI collaborators.
The advantage AI brings isn't mystical intelligence. One of its biggest advantages is much less exciting:
It's available.
A fleeting thought in my head can be captured and examined immediately. I don't have to ring up four engineers and ask if they have twenty minutes to discuss an AI CONSTITUTION for a project. I don't have to send an email and wait until tomorrow.
I can say, "Here's what I'm thinking."
And then I have to explain it.
That act alone makes the thought better.
A good collaborator can also tell me when my concept has the clarity of three-day-old lard.
That's useful.
Agreement isn't the goal.
4. Put your collaborator in the right chair
I have written technical books and articles with human editors.
Sometimes I had a copy editor.
Sometimes I had a technical editor.
Sometimes I had a prose editor.
They did different jobs.
An AI collaborator can sit in all of those chairs and more, but I have to tell it which chair I need occupied.
Sometimes I want my spelling and grammar checked.
Sometimes I want someone to challenge the thesis.
Sometimes I need research.
Sometimes I need a technical editor asking whether what I just wrote is actually true.
Sometimes I know exactly what I want to say and need help finding the words.
And sometimes I want my collaborator in the co-author chair.
Those aren't interchangeable requests.
"Write me 1,500 words about AI" is very different from "Here's an idea I've been wrestling with for three days. Help me figure out whether it holds water."
Know what kind of help you're asking for.
Then stay in the room.
Sandpaper smooths surfaces, properly used.
5. Make something
I build models.
And I hardly ever build anything straight from the box anymore.
The kit maker may have manufactured the plastic parts. They may have provided instructions showing where those parts are supposed to go. Someone designed the molds. Someone created the decals.
But when I look at that kit and form a story about it, something changes.
I cut open a door.
I make a moonroof.
I change the wheels.
I put the car in a scene that may or may not really exist.
The kit becomes mine as soon as I form a story about what I want it to become.
I think writing with AI is a lot like that.
An AI collaborator may produce some of the parts.
Sometimes it may produce a lot of them.
That contribution is real, and pretending otherwise would be silly.
But producing parts isn't the same thing as deciding what you're building.
The story matters.
The intention matters.
The choices matter.
And what you do with the parts matters.
6. Find yourself in what came back
This may be the part people skip.
Read the thing.
I mean really read it.
I read my writing aloud, both in my head and literally out loud. Years ago, I worked in radio, and somewhere along the way I developed what I call my microphone voice. Put a microphone in front of me and my rhythm changes a little. Things get a bit more polished. Ask me to say, "You're listening to the smooth jazz sounds..." and you'll hear what I mean.
That rhythm is useful when I edit.
I can hear when a sentence belongs to me.
I can also hear when something sounds weird.
I like fragments as pauses.
And I'll start a sentence with And.
But I'll also start one with But when the rhythm fits.
I don't particularly care whether the Oxford comma stays or goes. I can live with a dangling participle.
Spelling, however, needs to be correct, especially if I have decided to use some five-syllable word I've never used before.
I like made-up words when you immediately know what they mean.
What I don't like is language that makes the reader work simply because the writer wanted to sound clever.
If the substance of what I'm writing doesn't reveal my intelligence—or lack thereof—about a subject, fancy words aren't going to help.
Descriptive words are different.
"The humidity is 96%."
That's information.
"The air was as thick as old shoe leather this morning."
That's something you can feel.
Neither approach is universally better. The question is what the piece needs.
And sometimes what comes back from an AI collaborator is excellent prose that doesn't sound like me at all.
That's when my collaborator has another job:
Circle me back.
Don't manufacture a colloquialism. Don't sprinkle in fragments and pretend that makes something mine. Don't add a Southern expression because I grew up in Mississippi and Louisiana.
Tell me I've disappeared.
Then let me come back.
7. Put your name on it only when you own it
Here's where I think the argument about AI-assisted writing gets much simpler.
I asked the AI to write those sentences. And much like any documentation I write, I keep things as simple as can be. Small, short directives that are much easier to digest.
Then I read them.
Where necessary, I challenged them.
Sometimes I changed them.
Sometimes I threw them away.
Sometimes I read something it produced and thought, Yep. That's exactly what I was trying to say.
I am always in the loop.
If somebody asks an AI for 1,500 words, skims the result, and publishes it under their own name, the problem isn't that an AI generated the prose.
The problem is the skimming.
Engineers should recognize this immediately.
Imagine letting AI generate production code, glancing at it, and deploying it without understanding what it does.
That's dangerous.
And it's disingenuous.
The same responsibility applies to writing.
AI hallucinates. It can confidently create something wrong, irrelevant, unsupported, or just plain out there.
If it does that and I publish it under my name, I published it.
Authorship carries responsibility.
Understand what you're publishing.
Verify what matters.
Challenge what doesn't make sense.
Be able to defend the thesis.
If someone asks you why you wrote something, "the AI said so" isn't an answer.
8. Leave something behind
What do I hope survives after someone reads something I've written?
Maybe something changed in them.
Maybe they learned something.
Maybe they remember where they found it.
Sometimes prose is there to make technical information easier to remember. I've used story to teach engineers how to create Jenkins pipelines.
Sometimes it's there to preserve a memory.
My Dad always had a joke available.
He would have been delighted to know there were more people at his memorial "than a goat roping in downtown Jenner."
Someone who never met him might remember that line.
And if they do, for just a second they might have some idea of the man who said things like that.
That's writing doing its job.
Not every artifact lasts forever.
I wrote Applied jQuery. For about twelve seconds in 2011, it was useful.
Technology moved faster than me, the publisher, and probably everybody else involved. But for a little while, a couple of schools adopted it as a textbook.
It did its job.
Finished doesn't mean permanent.
Correct doesn't mean eternally correct.
Useful doesn't mean universally useful.
Write what you understand today. Tell the truth about what you find. Be willing to change tomorrow when the ground shifts underneath you.
AI doesn't change any of that.
It gives us another collaborator.
A remarkably available one. A sometimes brilliant one. A sometimes maddening one. One that can research, question, edit, organize, draft, challenge, and occasionally produce something with the clarity of three-day-old lard.
But the responsibility for the artifact carrying my name remains mine.
So maybe that old advice doesn't need much updating after all.
Write what you know or what you'd like to know.
Then stay in the room while you write it.
P.S. This guide is subject to change. Batteries not included. Some assembly required.