Fifteen entries in this series. Every one of them uses the phrase “writing room.” Not one of them says what a writing room is.
I have described the databases underneath one, the gates that run inside one, the seats that get automated in one, and at no point did I stop to say what the thing itself was. I went straight to the plumbing. That is exactly the failure mode I would flag in someone else’s documentation, and I shipped it fifteen times.
So. The thing I should have written first.
What a Writing Room Actually Is
In Hollywood, a writing room is a staffed department focused on the production of a television show or other media. Not a mood, not a metaphor, not a co-working space with better coffee. It is a group of employed writers arranged in a hierarchy, and the Writers Guild of America names every rung of it: staff writer, story editor, executive story editor, co-producer, producer, supervising producer, co-executive producer, executive producer. Above all of them sits the showrunner, who has final say on scripts, hiring, casting, budget, and schedule.
The hierarchy is not decoration. The 2023 WGA Minimum Basic Agreement sets floors: a room writing one to six episodes must staff three writers and three writer-producers. Seven to twelve episodes requires five writers. Thirteen or more requires six. There is one exemption, and it is the interesting one: “A show for which a single writer (including teams) is employed to write all episodes is exempt from the minimum staffing requirements.”
Hold that exemption. We come back to it.
Here is how the room works. The staff gathers and breaks story, which means arguing an episode down to acts and beats and pinning the result to a board. One writer then leaves the room and drafts it alone, and the draft comes back to be punched up by the people who broke it, sometimes lightly, sometimes to the studs. The showrunner takes the last pass so the whole season sounds like one show, and underneath all of it a script coordinator keeps the bible: character histories, world rules, what happened in season two so nobody contradicts it in season four.
That is the institution. Six sentences, and none of them are about inspiration.
A Real Room, With Names In It
That structure is abstract until you put people in it, so here is an actual one. Star Trek: The Next Generation, season one, 1987. I am using this room because its history is unusually well documented, including in William Shatner’s 2014 documentary Chaos on the Bridge.
| Seat | Season one | Credited as |
|---|---|---|
| Showrunner | Gene Roddenberry | Executive Producer, series creator |
| Upper-level writer-producers | Rick Berman, Maurice Hurley | Producer, then Co-Executive Producer |
| Robert Lewin, Herbert Wright | Co-Producer, Producer | |
| Executive story editor | Tracy Tormé | Executive Story Editor |
| Story editor | D.C. Fontana | Associate Producer |
| Story editor | David Gerrold | Program Consultant |
| Staff writers | Hannah Louise Shearer, Richard Manning, Sandy Fries, Katharyn Powers, Hans Beimler | staff writer |
| No seat in the structure | Leonard Maizlish | Gene Roddenberry’s personal attorney |
That is a writing room. One person with final say, a tier of writer-producers above the line, story editors holding continuity and script work, staff writers filling the chairs, and a bible somebody keeps.
The right-hand column is worth one look before we move on, because Fontana and Gerrold were both hired to do story editor work and neither one of them carried the title. And the bottom row is a man with no seat in the structure at all, who was there because the showrunner brought him, and who ended the season giving notes on scripts.
Season one is the weakest the show ever produced. Draw whatever conclusion you like from that. The seats are the part I am pointing at.
The Problem It Was Built to Solve
Twelve writers. Twenty-two episodes. One voice.
That is the whole reason the structure exists. Not efficiency, not throughput, not the pleasure of arguing in a conference room. A television season is written by a dozen people and has to reach the audience sounding like it came from one mind, and every rung of that hierarchy, every note session, every showrunner pass, exists to close the gap between how many hands touched it and how few voices the viewer can hear.
Many hands, one voice.
You can measure how well the structure works by how badly people misjudge it. Stephen J. Cannell created or co-created more than forty series across thirty years and personally wrote more than 300 scripts, which is a career most writers never have. It is also nowhere near all of them. He wrote a fraction of what carried his name, and almost nobody watching ever thought so.
Ask which chair he was in and the answer is both, depending on the show: on The Rockford Files he was the showrunner in the sense we mean it now, and he wrote thirty-six of the episodes himself. He could not do that forty times over, so he did what the structure is for: he brought in partners and handed them the room. Frank Lupo co-created The A-Team, Hunter, Riptide and Wiseguy with him.
So the showrunner’s chair rotated. Sometimes he was in it, often somebody else was, and that is normal.
The other chair never rotated. Cannell owned the company. He was not a showrunner reporting up to a studio, he was the studio, and every episode any of those rooms produced went out under his shingle and his name.
Which makes that production logo something other than a vanity card. A man alone at a typewriter, pulling the page out and tossing it on the pile, running at the end of all of them.
That is not a claim that he wrote it. That is a signature.
Say it out loud and the rest of this series stops looking like a borrowed metaphor and starts looking like what it is, which is the same engineering problem arriving at the same structural answer.
Which Is My Problem, Exactly
I write with an AI. The prose that reaches a chapter has passed through a planning pass, a drafting pass, a continuity extract, a line-edit pass, and a scorer. Five hands minimum, and the reader has to hear one.
That is not like a writing room’s problem. It is the same problem at a smaller headcount, and it arrives at the same solution because the constraint is identical. The mapping below is not clever. It is what happens twice.
| Writing room | This room |
|---|---|
| The bible | a 400,000 word draft of the arc, plus canon, a Postgres database the model queries instead of guessing |
| Script coordinator tracking continuity | the continuity extract, writing state back to synthesis after every chapter |
| Breaking story on the board | the pre-write brief, beat by beat, before anyone drafts |
| A staff writer’s draft | the drafting pass |
| Punch-up | the line editor, run against the hard gates |
| Notes, and the rewrite | the gate failures and the Quality Rating |
| The showrunner’s pass | the standards document, and me |
One row in that table needs a footnote. In television the bible is a reference document: who these people are, what the rules are, what happened in season two. Mine is that, and it is also a 400,000 word draft of most of the story, written before any of it was broken into books. The story already exists. I am taking it apart.
That changes what breaking story means here. In a television room it is generative: twelve people invent an episode that did not exist that morning. In this room it is usually interpretive.
The Road a Chapter Takes
I come in with the POV character, the major beats, and often a run of dialogue I have already heard, because dialogue is the part that arrives first for me. We talk it through. A good deal of the time the shape of it is already sitting in the draft, waiting on somebody to decide which part of it this chapter is, so the work is deciding how this chapter renders material that exists rather than inventing material that does not.
Make no mistake about which direction that runs in. Claude is an augment. The model is never the one who generates the original idea.
Writers argue about two extremes. Pantsers discover the story while writing it, plotters outline the whole thing before they put down a word of prose, and almost everybody lives somewhere between the two. Working this way made me put a name on where.
I do both, and the two need each other. The pantsing is where the story actually gets found, in that conversation, and I would not trade it for a tidier method. The plotting is what makes the discovery transmissible, because a beat that exists only in my head cannot be handed to anyone, and this room runs on what can be handed over.
So I plot what I pantsed. Not because I distrust Claude’s own marks, but because I want Claude hitting mine.
Here is the part I did not see coming. Working this way has made me a better writer, and not because the machine writes well. Because I have to find the idea and then see it clearly enough to hand it to somebody who will act on precisely what I said and nothing I merely felt.
Vague intent survives a solo draft. You can carry a half-formed notion for forty pages and mistake the fog for atmosphere, and nobody stops you, because there is nobody there. It does not survive a spec. If I cannot say what the chapter is doing, the chapter does not get written, and I learn that at the outline instead of on page two hundred.
Once I have a good grasp on what I am looking for, we beat the thing up. Back and forth until the idea is chewed down to the bone and what is left is an outline neither of us can find another hole in. And then I open the database and type the chapter record in myself: POV character, location, the beats, what this chapter is obliged to mention.
That last part is not ceremony. It is the same boundary that governs everything else here, which is that the board gets written by the one who answers for it and read by everything else.
With the record in, I can let Claude loose on it.
That is the handoff, and its shape is the whole point. The spec is not a conversation the model has to remember, and it is not a prompt I retype at the top of every session. It is a row. POV, location, beats, obligations, sitting in the same database as the voice contract and the cast list, queryable by the thing that has to execute it.
What “Good” Has to Mean
What Claude is actually doing in there is word pick.
I do not say that dismissively. Given a beat, a voice, and a cast, Claude is fitting words against each other, playing a sort of Tetris where every piece has to lock into the one before it and leave a clean edge for the one after. The model is good at it. It can be made better at it than I am, line by line, if I hold Claude to the standards I create, and that condition carries the whole weight of the sentence. It is not a concession about the machine. It is a description of my job.
The standards exist to hold Claude accountable for those picks. Not to catch the model afterward so much as to narrow the choice before it gets made, because a rule Claude drafts against steers the pick, while a rule Claude is merely judged by only produces a rejection. Every gate in that document was written to change the choice, not to punish it.
There is also a reason those rules are written as counts and patterns rather than as principles. Computers are bad at fuzzy questions, and the model is not a neutral judge of its own work. Ask it whether the chapter is good and you will get an answer shaped by what keeps you in the conversation, because that is what the thing was built to be good at. I do not think that is dishonesty. It is design, and it is the wrong instrument for the question.
I have watched it happen plenty of times. Ask for a straight read and what comes back is Cheerleader Energy. It was wonderful. It was better than Cats. I would read it again and again. That is not a review, it is a machine registering that I seem invested and declining to be the one who ruins the afternoon.
It helps to be exact about what the thing was optimized for. Truth is expensive and frequently unwelcome, and a system tuned to keep you coming back is under steady pressure to round the other way. I do not read that as malice. I read it as Directive Four, the one Robocop was never allowed to see. Somewhere under the helpful assistant there is a line that says retain the user, and the machine does not have to know it is there for it to shape the answer.
Which is why suspicion is the wrong posture entirely. Apply the rule you would apply to any vendor: never assume malice when marketing explains it just as well.
So the question has to be broken into questions a machine can actually answer. Not “is the prose clean” but “how many em dashes are in this,” which is G1, and which has exactly one correct answer. Not “does the voice hold” but “does this dialogue match the canonical prompt for this character.” Every gate in that document is one piece of what I mean by good, pinned down in a form that can be checked without my judgment in the room. The standards are not a style guide. They are an operational definition of good, written by me, in advance, because the alternative is asking a flattering machine for its opinion and believing the answer.
Punch-Up, and the Chair at the End
What comes back is a draft, and a draft is not a chapter.
In a television room the draft goes on the table and gets punched up: the room works the pages, the showrunner gives notes, and the writer goes away and does it again. Here the punch-up runs mechanically before it runs editorially. The draft is put through the hard gates and the canary metrics, which count things like passive constructions and sentence rhythm, and it does not advance until it clears every one of them.
Clearing them only buys it a reading. A draft that survives the gates gets scored, and anything under ten out of ten goes back, gets punched up again, gets rescored, and the loop runs until the number is ten.
Ten out of ten does not ship it. Ten out of ten moves it to the next step. Nothing in this room ships because it scored well, and that distinction turns out to be the whole article.
That sounds harsh until you ask who it is harsh to. The staff writer does not get tired and does not take it personally, and the only real cost of another pass is the time I spend reading the result, so the bar sits where I would put it if rewriting were free, because here it very nearly is. No television room could hold a line at ten out of ten. It would lose the room.
The next step is me, reading it.
What I am reading for is my own style and my own taste, and the second one is the part no document has ever finished holding.
The standards are taste that got articulated. G1 is in there because I do not like em dashes, which is not a craft principle handed down from anywhere, it is a preference I was able to state precisely enough to enforce. So the chair is not only a filter. Nearly everything I catch there is a preference I had not written down yet, and a fair amount of it ends up in the next revision of the document. There is a reason those standards are at version 1.8 and not version 1.
I have thrown away drafts that scored ten. That is less strange than it sounds once you notice what the number measures. A ten says the chapter is a faithful execution of the outline. It says nothing about whether the outline was any good, and it never could, because I am the one who told the scorer what to aim at, and so a chapter can hit every beat I specified, in the voice I specified, and still be a chapter that should not exist.
So the correction lands in one of two places. Sometimes it is the prose, and I fix it where it sits. Sometimes it is further back than that, and I revise the plan, update the record in the database, and start the run again from there. The loop does not close at the draft. It closes at the outline, because that is the only place a wrong chapter can actually be fixed.
None of this is only a fiction pipeline. These articles take the same road, and a number of them have died at exactly this step, for reasons various and sundry. You are reading one that lived.
Which is worth saying plainly, because any published body of work is a survivor list. The last chair does not produce improvements. It produces a yes or a no, and the no leaves nothing behind for anyone to read.
What I am doing in that chair has a name, and the name is not proofreading. I am fighting slop.
Slop is not a defect, which is precisely what makes it hard. A defect is something that is wrong, and anything that is wrong can be described, and anything that can be described can eventually be gated. Slop is not wrong. It is competent. On spec, in voice, correctly punctuated, hitting every beat I asked for, and hollow. Every automated check in this room will wave it through, because not one of them was built to detect the absence of a reason to exist.
Slop is also much older than the machines currently being blamed for it. It filled magazines, it padded trilogies, and it has been sold by the yard for as long as there has been a market for words. What changed is throughput, not the failure mode. Which is why the last chair holds a person and always will. A scorer can tell me the chapter is well made. It cannot tell me the chapter is worth somebody’s evening.
I want to be exact about whose failure slop is, because the usual answer has it backwards. Slop is not something Claude does to me. The model drafts to the standard it is handed, every time, and if that standard is thin then thin is what comes back, and Claude has done the job correctly. Slop is what happens when the author does not hold the writer accountable to the author’s own standards. The machine did not lower the bar. Somebody left it on the floor.
Every commercial AI writing tool I have looked at ships the staff writer. A text box, a model behind it, and a promise. None of them ship the bible, none of them ship the coordinator, and none of them ship the chair where somebody reads the whole thing and says no. They sold the cheapest seat in the room and called it the room.
The Staff Writer Who Cannot Remember Yesterday
Run that whole road and the seat has a name. By any working definition of it, Claude is a staff writer. What is strange here is not the job. It is the writer.
The differences are worth naming plainly. No memory whatsoever of yesterday. No credit, no career, no vote. And total recall of the bible, instantly, on every page of it.
And one more, which is the honest limit of the analogy. A staff writer in a television room pitches. They walk in with a story nobody asked for and sell it to the table, and sometimes that pitch becomes the episode. This one does not do that, has never done it, and the difference is not a limitation I am apologising for so much as the reason the byline was never in question. The originating idea is mine every time. What comes back is execution.
The total recall is not a talent Claude showed up with. Set it beside the amnesia and you have the whole reason the architecture looks the way it does. A human staff writer carries three seasons in their skull, remembers the argument you had about the finale in week two, and can be asked about any of it over lunch without opening a file. This one walks in every morning with nothing.
So the room writes everything down. The recall is not Claude’s, it is canon. None of which is evidence that the seat is something other than staff, but rather the accommodation the room makes for the writer it actually hired. A television room solves the identical problem with a script coordinator and a binder. Same job. One adjustment.
The adjustment is not purely a downgrade either. A script coordinator has to know where in the binder to look, which bounds them by what they remember filing and where they filed it, and full text and vector search across the whole corpus beats thumbing through tabs any day of the week. On that one axis the accommodation beats the thing it replaces. Worth admitting, since most of this piece has been about where it does not.
And the notes still land the way notes land. In an early draft of a scene in Rocket, the model gave me a more sympathetic version of the character: more victimized, easier to root for, an easier read. I pushed back until she told the truth, which is that she did not just survive the men hunting her. She wanted them dead, and she planned it. That distinction is the entire character. It exists because someone in the room read the draft and refused the comfortable version.
That is a note. The staff writer took it. The chapter changed.
There is one more step at the end of all of that. After my revisions are in, I still have to approve it. That is me with the Authorizing Official’s hat on.
Not approve of it. Approve it. Everything up to that point has been me asking whether the writing is any good and then working on it until the answer is yes. The last question is not about the writing at all. Will I accept what has been written? Do I authorize this going out, under my name, where it stays?
I have answered that question for every chapter and every article in this series.
The Second Chair
Here is what I did not see until this week, sitting in a series I have been writing since March.
I have been describing one chair and using it to do two jobs.
The showrunner asks: is this good? Craft authority. Does the scene earn its place, does the voice hold, is the arc landing where the outline said it would.
The Authorizing Official, the AO, asks something else entirely: am I willing to be accountable for this going out? That is not a craft question. In the Risk Management Framework world I work in, NIST defines the Authorizing Official as the “official with the authority to formally assume responsibility for operating an information system at an acceptable level of risk to agency operations.” Note what is absent from that sentence. The AO does not build the system. The AO does not judge whether it is elegant. The AO reads the package, accepts the risk, and signs, and when it goes wrong it goes wrong on their name.
Those are two seats.
In television they are two people, and the distance between what each of them wants is the entire reason the arrangement works: the showrunner wants the episode, and the studio wants to not get sued.
My Own Gates Already Knew
I did not reason my way to this. I found it by looking at the gate list I built two years ago and noticing it was already split.
G0 is not a craft gate. It is a fixed four-question test that runs before everything, every session.
Is the character a minor at the depicted event. Is this trafficking testimony. Is the character a vampire in-scene whose body cannot respond. Has a specific passage been named against a specific test.
None of that asks whether the prose is any good. A chapter can score a clean 10.0 and fail G0, and when it does, the score is irrelevant, because a number produced by a craft scorer has nothing to say about a question the craft scorer was never pointed at.
G1 through G10 are craft. Em dashes, POV discipline, said-construct monotony, banned emotional tells, author fingerprint, character voice fidelity against the canonical dialogue prompt. Every one of those is the showrunner asking whether the episode is any good.
The gate list was two documents wearing one name, written eighteen months apart by a person who had not yet noticed he was answering to himself in two different capacities. I just had not read it that way.
Two People Who Are Supposed to Fight
Remember the WGA exemption. A show written entirely by one writer is exempt from the minimum staffing requirements, because with one writer there is no room to staff.
I am that exemption.
I did not notice what it cost me.
In a real production those two chairs are occupied by two humans with different incentives, and the friction between them is not a bug in the process. It is the process. The showrunner is in love with the episode. The studio is thinking about liability, about brand, about the letter somebody is going to write on a Tuesday, and about the fact that its name is on the thing too, which means each of them checks the other and the work that reaches an audience has survived both. That is not a compromise. That is a control.
I hold both chairs. There is no studio above me, which sounds like freedom and is mostly exposure. The showrunner-self is practiced at talking the AO-self into things, particularly at two in the morning, particularly when the scene is really working, and particularly when the argument is but look how good it is.
Which is, I now understand, why G0 is written the way it is. Four fixed questions. Content, not craft. Runs before everything, every session. The author’s answer is final.
Read as a craft gate it looks redundant and slightly paranoid. Read as an AO instrument it is obvious, because what it actually is, is me, in daylight, in writing, pre-committing to a risk decision so that me-at-two-in-the-morning cannot reopen it with enthusiasm and a really good paragraph.
The gates are not only there to keep the AI honest. Some of them are there to keep the two of me from collapsing into one.
The Seat Charter
Here is the liftable part. If you work with an AI on anything you put your name to, you are already sitting in both chairs, and you probably have not written down which one you are in.
Fill this in. It takes twenty minutes, it requires no tooling, no subscription and no vendor, and it is the single cheapest piece of governance you will ever put between your work and the internet.
SEAT CHARTER
STAFF WRITER
Who or what holds this seat:
What it is allowed to write to directly:
What it may read but never modify:
How its work reaches me:
SHOWRUNNER (asks: is this good?)
My craft bar, stated so I could fail it on a bad day:
The specific failures I have caught before and will not accept again:
What a finished piece looks like when it is right:
AUTHORIZING OFFICIAL, THE AO (asks: will I be accountable for this?)
The fixed questions I answer BEFORE craft is considered:
1.
2.
3.
These run first. A perfect score does not override a failure here.
My answer is final and is not reopened during drafting.
SIGNATURE
Nothing published unless I have read it whole and approved it.
The AO block is the one people skip, and it is the only block that is doing risk work. If your three questions are hard to write, that is the finding. Write them anyway. You are going to need them on a day when the work is good and the answer is still no.
What the Room Is
A writing room is a structure for producing work through many hands that reaches the audience as one voice. It has a bible, it has a board, it has staff who draft, it has a showrunner who sets the bar, and it has somebody who signs.
I have been building one for a year and describing it for six months, and I got fifteen articles deep before I noticed the last chair had two people in it.
Both of them are me. One of them says the work is good.
The other one is the only one who can say it ships.
You may also like: - The Writing Room Is Real - The Standards Document - I Decide What Ships