Most of what gets sold as prompting advice is a promise that the right words in the box will hand you a finished book. They will not. If this series has shown one thing, it is that the book is won and lost in the structure, the honesty, and the nerve, none of which lives in the prompt. So before anything else, here is the honest ceiling: prompting controls how the machine drafts the surface and how well it obeys you. It does not write your book. It cannot supply a spine, pay off a stake, or find the courage to leave a wound open at the end. Those are yours. What good prompting does is make the drafting step clean and fast, so that the work you own, the deciding and the structuring and the revising, starts from a better place and wastes less of your time.
That is a smaller promise than the fantasy, and it is a real one. A writer who prompts well spends far less of the revision pass cleaning up tells, because the tells never got generated. A writer who prompts badly generates a wall of machine defaults and then fights it for a week. The difference between them is not talent or secret words. It is method, and the method is learnable in an afternoon.
I use AI every day, and it’s worth knowing what for, because it shapes this advice. It turns my interview transcripts into session summaries and roughs out first-pass outlines, and then I do the writing. I’ve got no patience for people selling prompt packs as a shortcut to a finished novel. They’re selling a fantasy to the writers who can least afford the wasted months. The prompting craft here is smaller than that pitch, and it works, and the examples come from how I talk to the tool.
Should you constrain or describe your voice when prompting AI?
The instinct most writers bring to a prompt is description. Write me a tense scene where Eleanor confronts the solicitor. That is a request for the machine’s average idea of a tense confrontation, and the machine’s average is exactly the hollow thing the rest of this series is about. Description gets you the default.
The prompt that works is mostly constraint. You are not telling the machine what to write so much as telling it what not to do and how to vary what it would otherwise make uniform. This is because the machine’s failures are predictable, and a predictable failure can be forbidden in advance. You know it will reach for inflated words, so you ban them. You know it will cluster every sentence at one medium length, so you order it to vary. You know it will have the character watch herself feel things, so you forbid it. You know it will end the scene on a soft note of significance, so you tell it not to. A good prompt reads less like a brief to a writer and more like a list of standing orders to a talented employee whose bad habits you already know cold.
I blame the tool vendors for the description habit. Their marketing tells writers to just describe what they want, as if the machine could read minds, and the writers who believe it end up with a manuscript full of averages they can’t sell. When you describe a scene to the machine, you’ve handed it every decision that matters, and it’ll make each one the ordinary way.
This is why the work of the diagnosis articles pays off here. Every tell you learned to see is a constraint you can write into a prompt. The banned word list becomes a line in the prompt. The rule against self-watching characters becomes a line. The instruction to vary sentence and paragraph and chapter length becomes three lines. You are pre-loading the machine with everything this series taught you, so it does not make the mistakes in the first place instead of making you clean them up after.
Show the machine, do not tell it
Adjectives about voice are nearly useless in a prompt. Write it in a spare, masculine, understated style produces the machine’s idea of those words, which is not your idea of them, and the gap between the two is where the draft goes wrong. The machine does not share your reference points. It shares the average of everyone’s.
The fix is the oldest rule in writing, turned on the tool: show, do not tell. Give it a passage in the voice you want and tell it to match that, and it will match the thing on the page far more closely than it will match any description of the thing. A paragraph of your own prose, or a paragraph by an author whose rhythm you are after, teaches the machine more in three sentences than a paragraph of adjectives teaches it in thirty. The example carries information that the description cannot, the actual cadence, the actual restraint, the actual way the sentences turn.
The same trick works in the negative. If the machine keeps producing a thing you do not want, paste the thing it produced and say, not this, and here is why, and then give it the version you would accept. Examples on both sides, the wanted and the unwanted, narrow the target faster than any amount of explanation. You are not describing the voice. You are demonstrating its borders.
I’d tell anyone to put their effort into the examples. Writers love fiddling with the wording of a request and hate the dull job of pasting in examples, and the dull job is the one that works. If you won’t show the machine what you want, you’ve got no grounds for surprise at what it hands back.
How do you build a style guide for AI?
The single most useful version of show, do not tell is not a paragraph dropped into one prompt. It is a permanent document that carries your voice into every prompt you ever write. Most writers never build it, and it is the thing that turns a generically clean draft into one that sounds like you and not like a competent stranger.
The distinction that makes it work is this, and it repays getting straight. The specification, the block this article keeps pointing at, is universal. It is the anti-default rules, the banned words, the no em dashes, the vary-your-lengths, and those are the same for every writer on earth, because they fight the machine’s defaults and every writer shares the same defaults. Your style guide is the opposite. It is the part that is only yours. It tells the machine not how to avoid sounding like a machine, but how to sound like you in particular, your diction, your rhythm, the things you do that another good writer would not. You need both. The specification makes the prose clean. The style guide makes it yours.
Building it is not mysterious. You assemble it out of your own work, because your own published prose is the most honest record of your voice there is.
Start with samples. Pull three or four passages from books you have already written, ones you are proud of, that sound the way you want to sound. Range them across modes, a passage of dialogue, a passage of description, a tense moment, a quiet one. These are the core of the guide. The machine learns more from four real paragraphs of yours than from any list of adjectives, because the paragraphs contain the thing the adjectives can only point at.
Then make the implicit explicit. Read your own samples as if you were editing a stranger and write down what you notice. Do you favor short sentences and let the long one land for emphasis? Do you avoid adverbs? Do you stay in the concrete and refuse abstraction? Do you let characters go unexplained? Do you use particular punctuation a certain way and never another? Whatever is true of your prose, name it as a rule. You are translating your instinct into instructions, because the machine cannot read your instinct but it can follow an instruction.
Then name the negatives, the things you would never write. This is as important as the positives. If you hate the word just, say so. If you never open a chapter with weather, say so. If you find rhetorical questions cheap, forbid them. Your style is defined as much by what you refuse as by what you reach for, and the machine needs the refusals spelled out because its defaults are full of exactly the moves you avoid.
The result is a document, a page or two, that you keep and reuse. It goes in at the top, the samples and the rules together, before the outline for the chapter. The machine then works against three things at once: the universal rules that keep the prose clean, the style guide that keeps it yours, and the outline that keeps it on the story. Like the specification, you build the style guide once and refine it slowly, and it stays the same from book to book because your voice does.
How long should a prompt be?
There is a temptation, once you understand constraint, to write one enormous prompt that tries to control everything at once and generate half the book in a single go. Resist it, for the same reason the drift article gave: the longer a single generation runs, the more the voice erodes back toward the machine’s defaults and the more the facts slip. A prompt that asks for one scene, against a clear specification, produces something you can hold to your standard. A prompt that asks for five chapters produces a slow landslide back into the average, and the landslide is worse the further it runs.
The one-giant-prompt approach gets sold as a productivity hack, and I think it’s some of the worst advice in this whole field. The writer saves an afternoon of prompting and then spends weeks untangling a manuscript whose voice slid back to the machine’s somewhere in the middle. Plenty of them never untangle it. They publish it, and their readers pay for the shortcut.
So you scope tight. One scene or one chapter at a time, never more, with the rules fresh at the top each time. You generate, you stop, you read, you fix, and only then do you generate the next piece, rules reset. Resetting the rules at the start of every generation is not redundancy. It is the thing that keeps the machine’s grip from loosening, because the machine does not carry your standing orders between generations the way you imagine it does. It carries a window, and the window erodes. Short prompts keep you inside the part of the window where the machine is strong.
Iterate. The first answer is a draft of the answer
A prompt is rarely one shot, and treating it as one is where a lot of frustration comes from. The machine gives you a first pass. You read it, you find the three places it slipped, and you push back, specifically. Not regenerate, which rolls the dice again, but a correction: this paragraph went purple, rewrite just this paragraph plainer. This line has her explaining what she already showed, cut the explanation. You keep the good and fix the local fault, the way you would direct a person, and the machine holds the correction far better than it invents a clean draft from scratch.
The mental shift is to stop thinking of the prompt as a vending-machine slot and start thinking of it as a conversation with a fast, fluent, slightly tone-deaf collaborator. You say a thing, you see what comes back, you adjust. The skill is not in the first prompt. It is in the second and the third, the corrections that walk the output toward what you want, one specific note at a time.
The specification is the prompt that matters most
Everything above comes together in one place, and it is the single most valuable prompt you will ever write: the specification. This is the standing block of instructions you put at the top of every generation, the reusable set of orders that tells the machine how to write, what to avoid, and how to vary what it would make uniform. It is the banned words and the no em dashes and the vary-your-lengths and the no self-watching and the let-dialogue-be-ragged, all of it, in one block, carried from chapter to chapter and from book to book.
You write it once. That is the fortunate part. The banned words are the same for every book. The orders about structure and dialogue and restraint are the same for every book. You build the specification one time, refine it as you learn what your particular model slips on, and then it is yours for everything you ever write. The only thing that changes book to book is the outline. The specification is where all the prompting craft in this article gets banked into a permanent asset, and the next article, on the full method, shows exactly how it sits at the top of the work.
The specification and the style guide are a pair. The specification is the universal half, the rules that keep any prose clean. The style guide is the personal half, the rules and samples that keep the prose yours. Together they are the standing context you hand the machine before every chapter, and between them they do almost all the work that prompting can do. Build both, keep both, reuse both.
Most writers skip this part, and I understand why. Writing a specification is tedious, and nobody applauds a banned-word list. A writer who won’t spend a weekend building one will pay for that choice on every book afterward, one cleanup pass at a time.
What can prompting buy, and what can it not?
Hold the honest line at the end, because it is the whole point. Good prompting gets you a cleaner surface, better obedience, less cleanup, and a draft that starts closer to your standard instead of further from it. That is genuinely worth having, and most writers never get it because they describe instead of constrain and they generate long instead of short.
What prompting can’t buy is the book. No prompt decides what your story’s about, what it costs or what shape it takes. Those stay in your hands however well you prompt, and the next article picks up there. Use the machine to get words on the page if it helps you. The people who make me angry are the ones selling prompting as authorship, because the writers who buy that pitch put their names on books they never made a single real decision about, and readers can feel it.
Next in the series: with the prompting craft in hand, the final article puts it to work. See how to write with AI, the full method that keeps the structure and the nerve yours while letting the machine draft. Or step back to the hub for the whole series, and see it all demonstrated in the annotated machine story and the rebuilt version.
