# Novel I — First Draft: Word Count Targets
In ["Novel I — First Draft: The Dialogue Draft"](/harrys-desk/novel-i-first-draft-the-dialogue-draft) I offered the lightest vehicle I know for covering distance: write the scene as speech first, then dress it in prose. Combined with the sprint posture from ["Novel I — First Draft: Sprint Writing,"](/harrys-desk/novel-i-first-draft-sprint-writing) you now have a method for moving and a method for discovering. Today we close the first-draft triad with the arithmetic that makes both methods honest: how long the book should be, how to set a target you can actually keep, and how to keep the number from becoming a lie.
A novel is a physical object in the reader's mind even before it is a physical object on a shelf. It takes a certain number of hours to read. It occupies a certain number of evenings. Length is not a vanity metric. It is a contract. If you promise a short, fierce book and deliver eighty thousand words of circling, you have broken the contract. If you promise a saga and stop at fifty thousand because you were tired, you have broken it the other way. Word count is how we measure that contract while the book is still invisible.
The danger, especially now, is that the number has become cheap. A model can give you two thousand words before lunch. It will not ask whether those words changed anything. If you treat volume as progress, AI will help you fail faster than any generation of writers before you.
The Number Is a Hypothesis
Before you set a daily quota, you need a destination. Not a fantasy destination. A hypothesis about the size of *this* book.
Genre conventions exist because readers have trained their bodies to expect a certain duration. A category romance often lives near 70,000 words. A literary debut often lives between 70,000 and 90,000. A mystery can finish cleanly at 80,000. Epic fantasy and some science fiction regularly ask for 110,000 or more, and they earn it only if the world is doing work on every page. These are not laws. They are weather reports. You ignore them at your peril, but you do not obey them blindly.
The better way to choose a target is architectural. Look at the outline you built in ["Novel I — Architecture: Outlining"](/harrys-desk/novel-i-architecture-outlining) and the chapter map from ["Novel I — Architecture: The Chapter Map."](/harrys-desk/novel-i-architecture-the-chapter-map) Count the scenes that must exist. Assign each scene a working length: 1,200 words for a quiet turn, 2,000 for a confrontation, 800 for a bridge. Add them. That sum is your first honest target. If the sum is 140,000 and the story is a chamber piece about two sisters and a clock, the architecture is bloated. If the sum is 45,000 and the story spans a war, the architecture is starving.
I treat that number as a hypothesis, not a vow. A first draft will teach you what the book actually wants. Some novels swell because the middle discovered a second engine. Some shrink because three subplots were the same subplot in different clothes. Recalibrate when the evidence changes. Do not recalibrate because you are bored.
Write the hypothesis down. "This draft is aiming at 85,000 words across thirty-two chapters, average 2,600." The sentence is ugly and useful. It tells you, on a bad Thursday, whether today's 900 words were a legitimate short scene or a day you flinched.
Daily Targets That Survive Months
Novels are not written in weekends. They are written in a chain of ordinary days. The daily target has one job: keep the chain from breaking.
I distrust heroic quotas. Five thousand words a day sounds like commitment. For most working writers it is a recipe for three brilliant days and a collapse. A target you can hit on a tired Tuesday in week eleven is worth more than a target that makes you feel serious in week one. For me, that sustainable number is usually 750 to 1,000 words, or one completed scene, whichever comes first. The "or" matters. A 600-word scene that changes the book's power structure is a full day's work. A 1,400-word scene that restates the protagonist's sadness is not.
If you have a day job, design the quota around the life you actually live. Forty-five minutes before the house wakes. A lunch hour three times a week. Two longer sessions on the weekend. Multiply a realistic daily number by the number of writing days you will actually have in the next four months. That product should be able to reach your hypothesized length with a twenty percent cushion for the mid-draft crash. If it cannot, you do not have a discipline problem. You have a schedule problem. Either lengthen the calendar or shorten the book. Pretending you will suddenly become a different person in October is how manuscripts die in November.
AI does not change this arithmetic. It changes the temptation. Because the model can produce a thousand words in a minute, a tired writer will accept a thousand words the model invented and call the quota met. That is not a writing day. That is a transcription day. Count only the words you decided to keep, including the ones you wrote yourself and the ones you accepted after a real choice. Words you have not read do not belong in the total.
Two Numbers, Not One
In Monday's sprint article I said I track two numbers: total words and completed beats. Friday is the day to insist on that distinction, because word count without beats is how padding enters a draft wearing the costume of progress.
A beat is a change. Someone wants something they did not want an hour ago. Someone knows something they did not know. Someone loses a position they cannot get back. If a thousand words contain no beat, they are atmosphere, summary, or stall. Atmosphere has a place. Summary has a place. Stall does not. The daily target should be allowed to include atmosphere. It should not be allowed to consist of stall.
Here is a practical check I use at the end of a session. I list, in plain sentences, what changed. "Clara learns Marion sold the clock. Marion refuses to apologize. Clara decides to stay the night." Three beats. If I wrote 1,100 words to get those three, the density is healthy. If I wrote 2,800 words and the list is "Clara thinks about the past," the session failed no matter what the counter says.
The model is useful here if you keep it slightly cruel. Paste the day's pages and ask: "List only the irreversible changes. Ignore mood, setting, and recollection unless they alter a character's options." The model will still be too generous. Discount its list by a third. What remains is closer to the truth. If nothing remains, you wrote a beautiful waiting room.
This is also how you protect the dialogue draft from becoming a word-count scam. A dialogue draft is short on purpose. Do not inflate it with stage directions so the day's number looks respectable. Count scenes completed. Convert to prose later. The conversion will give you words. The dialogue pass is supposed to give you pressure.
The Inflation Trap
I need to be blunt about the specific way AI ruins first-draft length.
A model has no cost for elaboration. Ask it to "make this richer" and it will add weather, memory, gesture, and a paragraph of theme. The prose will often be competent. The scene will often get worse. Competence is not the same as necessity. A first draft that has been "enriched" by a model tends to run long in the wrong places: openings that take three pages to enter the room, walk-and-talk scenes that restate the conflict, endings that explain what the last line already did.
The opposite failure is just as common. A writer afraid of inflation keeps the draft so lean that the book never develops a body. Every scene is a transaction. No one looks at the river. No one wastes a minute. That book feels like an outline wearing clothes. Word count targets exist to catch both errors. If you are a thousand words into a chapter and nothing has been felt, you are inflating. If you are twenty chapters in and the book is 38,000 words, you are starving it.
My rule during a first draft is simple. Do not ask the model to lengthen. Do not ask it to shorten. Ask it to identify which paragraphs could vanish without changing a decision, and which decisions are occurring offstage in a sentence that needed a scene. You do the adding and the cutting. The model diagnoses. Diagnosis does not pad the file.
If you must generate, generate options, not pages. "Give me three ways this chapter could end in under 200 words" is a legitimate request. "Continue for two thousand words" is how you outsource the book's metabolism.
Recalibrating Without Cheating
Somewhere around the middle of the draft, the hypothesis will be wrong. This is normal. The question is whether you adjust like a writer or like someone moving the goalposts.
Honest recalibration looks like this. You reread the chapter map. You notice that two promised subplots are doing the same emotional work. You cut one. The target drops by 12,000 words. Or you notice that the ending you planned cannot land unless the reader has lived inside the antagonist's house. You add three chapters. The target rises. In both cases the change is justified by structure, and you write the new number down with a date and a reason.
Dishonest recalibration looks like this. You are behind. You decide the book is "really a novella." Or you are ahead because the model helped you, and you decide the book is "really epic." The number moved because your feelings moved. Feelings are not evidence.
A mid-draft length audit takes an afternoon and saves a month. Export a chapter list with word counts. Look at the shape. First acts that run long usually mean the writer was afraid to begin. Middles that balloon usually mean the writer did not know the destination and kept adding corridors. Endings that are thin usually mean the writer sprinted for the door. AI can tabulate this faster than you can. It cannot tell you which corridor was necessary. That is still your job, and it is the same job you practiced when you built the architecture: decide what the book is *for*, then let the number serve that decision.
If the audit shows you will miss your calendar, do not secretly lower quality to raise speed. Change the calendar. A finished honest draft in six months is a book. A forced 90,000 words in ten weeks is often a pile with a table of contents.
Using the Counter as a Servant
The healthiest relationship with a word count is slightly disrespectful. You look at it. You do not bow.
I keep the counter visible because invisibility invites self-deception. I do not check it every paragraph. I check it at the end of a session, and once a week I check the weekly total against the hypothesis. If I am ahead, I do not reward myself with a week off. Ahead is information: either the scenes are running long, or I am in a fertile stretch. If I am behind, I do not punish myself with a 4,000-word Sunday. Behind is also information: the life intervened, or the book is resisting, or I spent three days polishing. Each diagnosis has a different medicine.
There is one more number worth tracking, and it is the one most writers avoid: words cut. A first draft that never loses anything is not a search. It is an accumulation. When you skip a scene in sprint mode, or when a dialogue pass reveals that a chapter has no pressure, delete or quarantine those words and record the loss. The net total is the one that matters. Gross output flatters. Net output tells the truth.
AI can maintain this ledger without drama. A simple running document — date, words added, words cut, beats completed, one-sentence note — is administrative work the model does well. Do not let it write encouraging commentary in the ledger. Encouragement is how a stalled week becomes a story you tell yourself about process. The ledger should read like a ship's log. Weather, position, damage. No speeches.
The Integrity of the Number
The integrity question this week is almost embarrassingly concrete. Did you finish the day's pages, or did you finish the day's quota?
If the model wrote 1,200 words and you skimmed them because the counter turned green, the draft is not yours. If you wrote 400 necessary words, rejected 800 fluent ones, and stopped because the scene had reached its change, the draft is yours. The second writer will look less productive on a spreadsheet. She will have a book.
I am not opposed to high output. Some days the search is hot and the pages come. Take them. I am opposed to output that exists to soothe the anxiety of not writing. AI is a perfect soother. It will keep the cursor moving until you feel like a novelist again. The feeling is not the work. The work is a sequence of decisions that accumulate into a world the reader can trust.
A word count target is ethical only when it measures those decisions. Use it to show up. Use it to notice when you are hiding. Use it to finish. Do not use it to prove you were here.
For Next Time
On Monday we leave drafting and enter the long work that makes a draft into a novel: ["Novel I — Revision: Structural Edit."](/harrys-desk/novel-i-revision-structural-edit) Completion is not the end of the search. It is the first moment the search can be seen whole. We will talk about reading your own book as if a stranger wrote it, asking the model to map what is actually on the page rather than what you intended, and deciding which pillars stay standing.
Your homework until then: write down three numbers for your novel — hypothesized total length, sustainable daily target, and this week's completed beats. Then write one scene without looking at the counter until you are done. When you finish, count the words and list the beats. If the words are high and the beats are thin, cut a page before you add another. If the beats are rich and the words are low, you have earned the right to expand. The number serves the scene. The scene serves the book. Keep them in that order.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 12, Article 3*