The Work That Writing Tools Don’t Do
This post is part of my Medium blog.
The first book I wrote was Jakarta Commons Cookbook, published by O’Reilly in 2004. I remember the Word template better than I remember most of the chapters. You wrote in Word. You followed the template. You sent the manuscript off and then waited for somebody else to find the mistakes you could no longer see after reading the same paragraph forty times.
There was nothing automated about it.
That was book publishing twenty years ago, and I want to be precise about which part of it was hard. It was not the thinking. I liked the thinking. The hard part was everything stacked around it.
Proofreading. Testing the examples. Testing them again after the library changed underneath you. Hunting for the sentence where a word had gone missing. Reading a page until it stopped being language and became a pattern of marks, then reading it again anyway, because the sentence had to be right the first time a stranger hit it.
I have written or co-written four technical books. They took evenings, weekends, and whatever attention was left after dinner. None of that time went where people assume it went.
I’m dyslexic. I also have what I’m fairly sure is dysgraphia, and I’m not going to turn a blog post into a medical file. The practical effect is simple. I can write a sentence that is complete in my head and also manage to leave a necessary piece of it on the floor. A word disappears. A verb ending fails to arrive. The pronoun is missing.
This was before Grammarly and long before generative AI. Grammar checkers existed, but they were all pretty awful. The ones I tried in those ancient times all objected to things that were fine and stayed quiet about the things that weren’t. So the proofreading tool was me, reading the same paragraph until I could no longer tell whether it said what I meant.
Writing used to be a grueling exercise of proofreading, correcting, reviewing red line edits.
Those books taught me how to write a book. They also taught me that most of the job had nothing to do with having something to say.
After O’Reilly I kept writing. Some part-time journalism, open-source books for a startup that gave them away, whatever technical material the job required. But for thirteen or fourteen years I didn’t publish another book, and it wasn’t because I ran out of ideas. I knew what the surrounding work cost, and I wasn’t willing to pay it again.
That is the part missing from most arguments about writing tools. The barrier isn’t always talent or discipline. Sometimes it’s the specific tax the tools charge you, and back in 2004 it was a heavy price to pay that was measured in months.
About a year ago I started writing seriously again. Part of that is the book I’m working on now. Part of it is remembering that I enjoy writing now that writing isn't as focused on proofreading as it was in the pre-Grammarly, pre-Generative AI age.
The tools are different now, though not in the way people usually mean.
I don’t use generative AI to write my books line by line. I don’t prompt it and accept the answer as prose. The output is functional the way a temporary wall is functional: it keeps the weather out while you work out what the building is supposed to be. But, it is also not enjoyable to read.
Generative AI writes awful English. Even the most expensive models still output slop.
So I rewrite it. Every line that matters. Sometimes I type the replacement, sometimes I dictate it, and sometimes I keep the idea and throw out every word the model used to get there.
A functional first draft is not writing. It’s material to argue with.
The genuinely useful part starts earlier than drafting. I can hand an idea to an agent and send it after the surrounding facts. If I want a reference to Enrico Fermi, I can ask what he was doing at Hanford, who was in the room with him, and which part of that story actually matters to the point I’m making. The difference is that I can send the first pass away and keep holding the argument while it runs. I'm not spending nights and weekends researching facts or proofreading Word templates.
Some people hear that and call it a "shortcut." Google was a "shortcut." The public library was a "shortcut." Nobody asks whether a reporter at The Atlantic is committed to the work because they didn’t crank through the microfilm personally. The questions that matter are whether the research is accurate, whether the writer understood it, and whether the reference belongs in the piece at all.
The tool can find the material. It cannot decide what the material means here.
It’s useful for ideation too, in a way that has nothing to do with prose. I can describe a problem badly, ask for a few possible structures, reject most of them, and find out what I was actually trying to say in the process. That isn’t the model writing the essay. It’s closer to talking through a half-formed idea with somebody who has read too much and might have helpful feedback. I'm frequently capturing a stream of conciousness read out of a scene and then using Generative AI to discuss the plausibility of plot twists.
Eventually I might ask for a functional first draft, since starting from an empty page is not always a good use of the hours I have. But I don’t confuse that draft with the work. The work is the judgment that comes after it: which sentence is true, which one is merely plausible, which idea is mine, and which one showed up because the model knows how essays are supposed to sound.
That last category is the dangerous one. Models are very good at producing sentences that sound finished before they have said anything.
What the tools removed was the work that kept me out of the chair. Rereading a paragraph for the tenth time because I can’t see the hole in it. Chasing the same background fact while trying to hold an argument in my head. Forcing a structure out of nothing when I already know the problem but haven’t found its shape.
They did not remove the writing. They moved the rest of it aside so I could get to the part I wanted: deciding what I mean, saying it so a stranger can follow it, and giving that stranger a reason to keep going.
That is a different claim from “AI can write a book.”
AI can produce a book-shaped object. It can fill the pages, imitate the sound of a conclusion, and make a manuscript look as though it already survived the hard part.
It hasn’t.
The hard part is knowing a sentence is wrong even though it sounds good. It’s cutting the paragraph you spent three hours on because it doesn’t belong. It’s putting your name on a book once you are comfortable it isn't slop.
I’m writing books again because I missed that work. What the tools cleared away was the work I never wanted.
image: /blog/images/the-work-that-writing-tools-dont-do.jpg image_alt: "The Work That Writing Tools Don’t Do"
In Redundant, the first book in The Condition Set trilogy, Rob Coleman runs the FinOps review that names the waste nobody wants to hear about. The numbers don't change. The question is who they get used against.
image: /blog/images/the-work-that-writing-tools-dont-do.jpg image_alt: "The Work That Writing Tools Don’t Do"