Every scribblers’ room I’ve ever sat in has a version of the same wall. Index cards in different colors for different characters. String sometimes, if the mythology runs weird enough that somebody needs to see the connections instead of just say them out loud.
Doesn’t matter if it’s Alias or Lost or Heroes or a pilot that never got picked up. The wall is the room’s memory, because nobody at that table, no matter how good the pitching gets on a given Tuesday, can hold an entire season in their head at once.
I need that wall more than most people at the table. I’ve got ADD, diagnosed late, well into my career, after my son got tested for it and I recognized every symptom on his list as my own. Craft doesn’t stick in my working memory the way it does for most scribblers.
So a few years back, on a hiatus between shows, I sat down and built myself a personal version of that wall. Every fundamental I’d ever learned about scribbling a scene, scribbled out by hand: track the character emotionally, make it cinematic, the power of the time cut. A mnemonic I rewrote every morning until it stuck.
I wasn’t inventing anything. I was just doing by hand what a whole category of software wants to sell scribblers as a product now.
Two Volcano Movies, an Asteroid and a Comet
Hollywood’s had this problem since before I got in the business. Dante’s Peak and Volcano hit theaters the same year, months apart. Armageddon and Deep Impact hit the same summer after that, an asteroid picture and a comet picture.
Every so often the town develops a shared unconscious, and two studios greenlight the identical high concept in different buildings, neither aware the other exists ‘til the trades break it.
Something close to that just happened in AI scribbling software, except it happened three times to the same word.
Storyflow.so is a visual workspace, notes and research and story material and images sitting on a canvas the model can read. Storyflow-ai.com takes you from script to storyboard to AI-generated video.
StoryFlow Editor, out of a small studio called Soulstices, is a branching-narrative scripting tool that landed in Steam Early Access this past November and plugs straight into Unity, Unreal, and Godot.
Three companies. Three different jobs. Funny collision.
Ask why it happened and you land somewhere more useful than the joke. An entire industry asked a language model for something like story plus workflow, and everybody’s model handed back the same portmanteau.
We spent the first couple years of this technology arguing about which chatbot to use. That argument’s basically over. The one that matters now is what gets to walk into the room with actual authority once the model’s done proposing things.
What the Room Actually Does
BookBub surveyed something like twelve hundred authors last year and found about 45% using generative AI somewhere in the process. Skews hard toward working commercial self-published scribblers, so don’t read that as gospel for every scribbler alive.
But look at what the AI-using half actually does with it. Eighty-one percent for research. Outlining, marketing copy, organizing a plot, a second opinion on a scene that isn’t working. ChatGPT led at 85%, Claude a distant second at 54%.
A newer survey out of Kit this year, 550 working creators, found nearly the same shape: ChatGPT 73%, Claude 70%, Gemini 40%, people running two or three at once instead of picking one and going steady. And 89.2% of Kit’s respondents said they always review and edit the output before it goes near a reader.
Not one respondent claimed they used it straight, no notes. That’s every table read I’ve ever sat in. The room generates. Somebody with actual standing still calls it, works or doesn’t, before anybody downstream trusts it.
Sudowrite’s Story Bible is close kin to the mnemonic I built by hand on that hiatus, except it lives in software and the model can read it directly: characters, world rules, synopsis, beats.
Even Sudowrite’s own advice, once you’ve got something screenplay-shaped out of it, is to walk it over to Final Draft, Highland, or WriterSolo and let dedicated software handle the formatting a reader actually trusts.
Smart split. You want the guessing half of the stack good at guessing, and the counting half never guessing at all.
Novelcrafter takes the split further and makes it explicit. Its Codex is a project wiki, characters and places and lore, and you can point almost any brain at it: GPT, Claude, Gemini, Llama, Mistral, or a few hundred more through OpenRouter.
The story bible stays put. Whatever’s plugged into it is a rental. And Final Draft, sitting at the far end of that pipeline doing the least glamorous job in the building, still runs on something called a Beat Board. A digital corkboard. Software finally caught up to the wall.
The Guild Already Drew the Line
I’ve carried a WGA card since 2001. The 2026 MBA keeps the AI language the Guild fought for in the 2023 strike, and it’s worth restating, because people outside the business don’t understand how deliberately that language sits.
AI is not a writer under our contract. Material a company hands a scribbler that came out of a model doesn’t count as source material when the Guild decides whose name goes on screen.
A studio can’t make using it a condition of employment. If a scribbler wants to use it anyway, the company must consent, and whatever policy company has still governs on top of the contract.
The new agreement adds a piece I like. If a studio licenses our scripts, our life’s actual work, to train a commercial model that generates new material, the company must notify the WGA. The Guild can then sit down and talk about what scribblers are owed for it.
The Dead King Walks Into the Tavern
Games don’t get the luxury of fading out. A script can sit wrong for a week until somebody catches it in the next draft. A character inside a live game might be talking to somebody at two in the morning and has to remember, mid-conversation, who’s dead, which door is actually locked, what that particular player already knows and hasn’t been told twice.
Inworld built its name on exactly that problem, AI characters and NPCs. By this year the company describes itself several floors further down: realtime voice, speech recognition, model routing, context management, an orchestration layer moving across more than two hundred models from a pile of different providers. They’ve raised close to $120 million chasing a question that used to just be a narrative designer’s headache and is now infrastructure.
I’ve never had an AI character forget the rules on VALORANT. We don’t run generative dialogue there. But I’ve had a human scribbler do it. Pitched a scene deep in an Alias room once, built around a piece of Rambaldi tech our mythology said didn’t exist on our timeline. Happens in every room that runs long enough. The difference is somebody catches it before the table read.
That’s the company LoreWeaver is chasing, and their pitch line is blunter than most of what crosses my desk: large language models are fluent, not faithful. Their examples read like a narrative designer’s blackout nightmare.
A king you killed three scenes back wanders into the tavern and starts making conversation.
A scholar scribbled gentle as a single adjective in a design doc threatens somebody with violence because the sentence sounded plausible in the moment and nothing in the system stopped it.
Their fix sits a validation layer between what the model wants to say and what a player actually hears: pull the real canon, generate inside it, check the output against the rules, kill whatever breaks them. The model proposes. Your rules dispose.
The Note About the Note About the Note
The failure that actually worries me isn’t the dead king. You catch that one. It’s loud, it announces itself. It’s the slow leak.
Every scribbler who’s worked inside a network system knows this move already. A note comes down from the network. The head of current passes it to the exec assigned to your show, with a little spin of their own added on. That exec passes it to your showrunner, more spin. By the time it lands on you in the room, three people have already reshaped it.
A note about a note about a note, and every single person in that chain swears they’re just relaying what the last person told them.
AI scribbling systems are building the identical chain into their own architecture, on purpose. No model holds an entire long project in its context window at once.
Long projects lean on summaries, retrieval, character memories, compressed versions of scenes that happened three hundred pages back. The model generates something a little bit off. A second pass summarizes it. A third system retrieves that summary later and treats it as fact.
Nobody flags the original mistake three hops downstream. By then it doesn’t look like an error. It looks like established character.
Four Jobs, New Hardware
Strip the product names off everything above and you get the same four jobs every show I’ve ever run already had, just running on different hardware.
Somebody generates. That’s the room, or now, on a slow morning, a model sitting in for the room.
Somebody remembers. That’s the bible: a whiteboard, a Codex, a Beat Board, or a mnemonic scribbled out by hand on a hiatus nobody but me remembers.
Somebody turns pages into a thing that can actually get shot, printed, or shipped.
And somebody has to decide if the scene actually works, if the character would really say that, if breaking your own canon is a mistake or the best idea anybody’s had all week.
Adobe ran a survey this past May, better than sixteen thousand creators worldwide. Seventy-five percent called AI integrated or essential to how they already work. Eighty-five percent said the final creative call should always stay with the human being.
Read those two numbers side by side and they’re not fighting each other. The second one is probably what makes the first one survivable. Nobody hands over the wheel unless they trust there’s still a hand on it.
Guard the Wall
So this is where I’ve landed, and it’s a smaller point than the software companies are selling it as. The wall doesn’t go away because the pushpins can argue with you now. If anything it matters more, because it’s holding against something that can generate a thousand plausible index cards a minute and has no idea, none, which ones are true.
Pick your model. Swap it out next year. I’m not precious about ChatGPT versus Claude versus whatever ships next, and neither should you be.
But build the wall first, in your own hand if that’s what it takes, before you let anything start pinning cards to it for you.